SuperUROP All SuperUROP Projects ()
spacing
SuperUROP
 
Home
Letters of Recommendation
Login

SuperUROP Projects

"Towards an AI policy toolkit"
Faculty Advisor: Aleksander Madry
Mentor(s): Sarah Cen, Andrew Ilyas
Contact e-mail: shcen@mit.edu
Research Area(s): AI Policy
Data-driven algorithms often have harmful unintended side effects. For example, algorithm used to allocate healthcare resources unintentionally favor patients with a particular ethnicity because the algorithm designers overlooked that historically, patients from that ethnicity received more care for the same conditions. Because it is impossible to design technical fixes for every such corner case, *we must have laws* to protect individuals that fall through the cracks. Unfortunately, there remains a large gap between those with the technical expertise necessary to understand data-driven algorithms, and those with the political and legal expertise necessary to regulate them.

In this SuperUROP, you will help develop tools and resources towards bridging this gap. Your main responsibilities will be reading, summarizing, and synthesizing ideas from multidisciplinary texts. You should therefore be comfortable grasping the high-level ideas of, for example, both law review papers as well as ML papers. You should also have strong written communication skills.

Above all, you should be interested in the impact of AI and its governance. Please reach out if this describes you!
Towards an AI policy toolkit
"Economic and societal impact of multilayered AI-deployment"
Faculty Advisor: Aleksander Madry/Luis Videgaray
Mentor(s):
Contact e-mail: luisvc@mit.edu
Research Area(s): AI Policy
Increasingly, organizations utilizing machine learning (ML) outsource the creation of data and the training of the ML models. This results in a progressing vertical segmentation of the AI supply chain and, more broadly, the prevalence of the AI-as-a-Service model. In fact, ML delivery starts to constitute a chain of multiple businesses that ultimately connect to the end consumer. This trend tends to accelerate with the recent emergence of very large "core" or "foundation" models. This project will address some of the relevant economic and societal questions raised by the emerging configuration of AI deployment, such as:

* Who is accountable in an increasingly fragmented AI supply chain? How to tackle the risks that might arise at the "seams"? Who and wow will manage aggregated risks along the whole supply chain?

* Will AI multi-layered supply chains lead to increased concentration by only a few companies able to produce the large base models? What policies or regulations may introduce more competition? What is the optimal regulatory framework under an increasingly concentrated market structure at the top of the AI supply chain? Should base models (and their developers) be deemed essential facilities or systemically relevant for regulatory purposes? Would entity-based regulation be an appropriate paradigm?

* What are the geopolitical consequences of global supply chains fragmented across national borders? How should international trade of AI services be regulated? What are the opportunities and risks for countries that only participate in downstream segments of the global AI supply chains and what policies and regulations may increase their agency?
Economic and societal impact of multilayered AI-deployment
"Formal verification of distributed/concurrent Go code"
Faculty Advisor: Nickolai Zeldovich
Mentor(s):
Contact e-mail: nickolai@csail.mit.edu
Research Area(s): Systems (incl OS, databases, computer security)
Writing correct distributed systems software is challenging because, in a
distributed system, individual machines can fail and recover while others
continue working, and because different machines execute concurrently.
As a result, subtle and easy-to-miss bugs can lead to serious problems.

This project's goal is to explore the use of formal verification
to develop, specify, and verify correctness of distributed systems
implementations and applications.

The project will build on prior work in the PDOS research group, such
as the Perennial framework Link for
reasoning about concurrency and crash-safety using concurrent separation
logic and the Goose tool for lifting Go code into Perennial.

Possible project directions include extending Goose to support additional
features of the Go language (such as interfaces or channels), developing
verified libraries for distributed systems (such as marshalling or RPC),
or helping develop a verified distributed transaction system.
Formal verification of distributed/concurrent Go code
"Large Scale Production of Low-Cost Lightweight Homes using Additive Manufacturing of Recycled Polymers"
Faculty Advisor: David Hardt
Mentor(s): none
Contact e-mail: hardt@mit.edu
Research Area(s):
This project seeks a highly scalable solution to the problem of affordable, dignified homes for the unhoused and underhoused populations of the world. We propose to do this by creating very light weight, structurally sound homes using recycled polymers with additive manufacturing as the dominant method. The research includes design and structural analysis of typical homes, manufacturing process and factory design, as well as methods for factory siting and supply chain management.
"Simultaneous imaging of label-free autofluorophores for cancer diagnosis"
Faculty Advisor: Sixian You
Mentor(s):
Contact e-mail: sixian@mit.edu
Research Area(s): Applied Physics, BioEECS
Label-free imaging eliminates the need of stains or genetically modified indicators and has strong potential for translating to clinical diagnosis. Simultaneous imaging of label-free contrast makes this even more attractive because it reduces acquisition time and motion artifacts. In this project, the student will learn the principle of label-free imaging, learn to use multiphoton microscope to characterize label-free contrast, and investigate the optimal condition for simultaneous label-free imaging for cancer diagnosis. Simultaneous imaging of label-free autofluorophores for cancer diagnosis
"Fairness in computer vision"
Faculty Advisor: Lalana Kagal
Mentor(s): Schrasing Tong
Contact e-mail: lkagal@csail.mit.edu
Research Area(s): Cognitive AI, Computer Graphics and Vision, Human Computer Interaction, Machine Learning
Recent research has shown that trained models could produce biased predictions and discriminate against socially vulnerable groups. To address these problems, researchers have proposed various definitions of fairness as desirable forms of equality, often relying on sensitive attributes such as race or gender. However, evaluating whether a model is fair in the image domain remains extremely challenging for two reasons: (i) images often lack explicit labels on sensitive attributes and (ii) bias can arise from a much larger set of proxy features present in the scene, as opposed to only the defined features in a tabular dataset.

We propose a novel bias detection approach that uses an off-the-shelf object recognition algorithm to extract human understandable concepts and then train an interpretable proxy model to perform the classification. Although estimating the degree of bias may still be difficult, this approach scales well to real world datasets and serves as a starting point to bias mitigation by identifying possible features that are the root cause of bias. Join us in developing this methodology and contributing to a conference paper!

Some experience/courses in ML is required.

[The image is an example of racial bias in COMPAS, which has been used to predict which criminals are most likely to reoffend. ProPublica compared COMPAS's risk assessments for 7,000 people arrested in a Florida county with how often they reoffended (Angwin et al; 2016; Garber, 2016; Liptak, 2017) and found that black offenders were seen almost twice as likely as white offenders to be labeled a higher risk but not actually re-offend.]
Fairness in computer vision
"Visualizing and Predicting Human Behavior"
Faculty Advisor: Richard Fletcher
Mentor(s):
Contact e-mail: fletcher@media.mit.edu
Research Area(s): BioEECS, Computer Systems, Human Computer Interaction, Machine Learning, Signals and Systems
Understanding and predicting human behavior is one of the most important open problems in medicine and health care. The largest burden of disease worldwide (e.g. cardiovascular, diabetes, COPD, drug addiction, etc.) is now primarily due to our behavior and psychosocial factors. Our group has been developing a digital phenotyping platform that collects sensor data from mobile phones to help track and predict human behavior and mental health.
This SuperUROP project has two components: (1) to help design and create visual representions of human behavior (as indicated by phone activity and phone sensor data) that is implemented on our central server front end; (2) Using the collected data, to apply machine learning and build CNN and RNN (LSTM) models to predict specific behaviors and changes in mental health states. These two components could be done together or implemented as separate SuperUROP projects, depending on the level of student interest and experience.
Experience with machine learning and strong coding skills (Python, Javascript) is required. Experience with back-end/front-end server development and/or Android/JAVA coding is also a plus.
Dr. Fletcher also holds a research faculty appointment at Mass General Hospital, Dept of Psychiatry and is a 10-year member of Society of Behavior Medicine.
Visualizing and Predicting Human Behavior
"Let's Work Together...or Not!"
Faculty Advisor: Una May OReilly
Mentor(s): Stephan Moskal, Erik Hemberg
Contact e-mail: unamay@csail.mit.edu
Research Area(s): Cognitive AI, Computer Networks, Machine Learning
What are the roles of competition and cooperation in securingcomputer networks? How does competition among threat actors or network security personnel impede or improve the likelihood of success? To what extent can intelligent cooperation among different stakeholders in the defensive cyber ecosystem mitigate risk and damage? What are useful forms of cooperation and why are they harder or easier to encourage? What game theory or evolutionary cooperation theories are at play?

This project entails simulating agents capable of learning (via different machine learning techniques) and cooperating (via direct or implicit communication) in the context of abstracted cyber networks.
Let's Work Together...or Not!
"Curation and Intelligent Inquiry of Climate Change Denial Documents"
Faculty Advisor: Una May OREILLY
Mentor(s): Aruna Sankaranarayanan
Contact e-mail: unamay@csail.mit.edu
Research Area(s): Environment, Human Computer Interaction, Machine Learning, Natural Language and Speech Processing
The climate change movement is gaining momentum, and scientists, activists and
organizations alike are trying to persuade the masses about the realities of
climate change. In this context, a missing piece is the absence of an aggregated
database that compiles documents from climate contrarian organizations - fossil
fuel corporations and vested interests - who dispute the realities of climate
change, alongside scientific documents from climate change scientists. Current
methods for investigating such documents involve manual labor, and use existing
tools to query various sources on the internet. This project will kickoff the
creation of the first database on communications from the climate change denial
movement. It involves document curation and NLP techniques that simplify the process
of investigating climate change denial.

This project is ideal if you want to improve your knowledge of data science, natural
language processing with machine learning models, and interface design.

Figure: Taro Istok, CC BY-SA 4.0 <Link via Wikimedia Commons
Curation and Intelligent  Inquiry of Climate Change Denial Documents
"MuseScore Plugin"
Faculty Advisor: Garo Saraydarian
Mentor(s):
Contact e-mail: garo@mit.edu
Research Area(s):
This project consists of development of a MuseScore plugin that enables Musescore files to be created for musical dictation training that:
1. Attaches a sound file.
2.Converts solfege keystroke entry into standard notation and returns corrected version.
3. Separates rhythmic, melodic, and harmonic entry.
4. Provides individual user data such as number of corrections, most frequently missed element, etc.
MuseScore Plugin
"Reverse Engineering Shakespeare: Using Programming for Pedagogy and Performance"
Faculty Advisor: : Diana Henderson
Mentor(s): none
Contact e-mail: dianah@mit.edu
Research Area(s):
This project aims to "reverse engineer" drama, using Shakespeare's plays as corpus, to reanimate language written for embodied performance across media. Using insights from at-scale work in the digital humanities will provide fresh means of engaging dramatic texts through character and place. One major strand focuses on 'scenic building blocks' to reveal how dramatic characters work, while the other will re-map the plays' imagined locations in their historical, geographical, political, and mythic complexity. Centrally, our scenic building blocks strand will leverage the "study clips" functionality of our "Global Shakespeares in Performance" curricular modules, while 'mapping Shakespeare's world' will work from the Shakespeare texts and the curricular modules to articulate the plays' imaginative geographies.

SuperUROPs will organize and annotate dramatic locations through a limited, layered set of categories; use R and Python to process dramatic scenes as data; and synthesize results to create heuristics that help actors, directors, designers, audiences and researchers consider meaning in patterns of geography, space, and time.

Tools created will complement and interoperate with existing open-access resources in Shakespeare Studies. Using contemporary digital methods to revisit dramatic texts will provide rich possibilities for interacting with Shakespeare in performance, while leveraging new media affordances for future productions.
"Designing regulator cocktails to reverse brain aging"
Faculty Advisor: Manolis Kellis
Mentor(s): Alex Lenail
Contact e-mail: kellis-admin@mit.edu
Research Area(s): BioEECS, Computational Biology, Machine Learning
Our neurons degenerate as we age, showing protein aggregates, mitochondria failure, and epigenome erosion, and neuronal pruning. This has generally thought to be an irreversible process, but the groundbreaking experiments by Shinya Yamanaka that created pluripotent stem cells from skin cells gives us hope that carefully-selected combinations of transcription factors may reverse neuron degeneration. In this project, you will help identify such regulator combinations for reversing brain aging, by integrating large datasets of single-cell gene expression and epigenome accessibility (scATAC-seq, scRNA-seq), inferring regulatory networks and circuits, and predicting driver regulators of neurodevelopment and neural homeostasis for manipulation.

This project is ideal if you like programming, data mining, analyzing cutting-edge genomics data, and interpreting the biological results. We'll read and critique papers together, ask questions about the transcription factor biology and neural homeostasis, and see what holds up in the data. We'll design machine learning models to tease signal out from noise, and rigorously benchmark model performance. We'll craft custom visualizations for these data. Ideally, we'll discover some candidate regenerative therapies which collaborators will test experimentally in mice.
Designing regulator cocktails to reverse brain aging
"Deep Learning for Imaging-Genomics-Transcriptomics Single-cell Integration"
Faculty Advisor: Manolis Kellis
Mentor(s): Brad Ruzicka
Contact e-mail: kellis-admin@mit.edu
Research Area(s): BioEECS, Cognitive AI, Computational Biology, Machine Learning
The MIT CompBio lab (compbio.mit.edu) headed by Prof. Manolis Kellis is looking for talented SuperUROP students for computational projects in spatial-transcriptomics integration of imaging data in heart and brain, single-cell genomics, epigenomics, and disease circuitry dissection of brain disorders. Expertise in programming, data analysis, method development, and machine learning will be great strengths for a successful project. We have generated 2.5 million genome-wide single-cell maps across genetic, epigenomic, and transcriptional profiles across post-mortem brain samples from brain. We have already begun analyzing the data across Schizophrenia, Alhzeimer's, ALS, and other neurodegenerative and psychiatric disorders.

The goal of this SuperUROP project is to help analyze these single-cell datasets in the context of spatial transcriptomics information, and genetic variation from whole-genome sequencing, and phenotypic variation from rich longitudinal profiling and cognitive evaluations, enabling us to discover variants, regulatory regions, genes, pathways, cell types, and brain regions with causal roles in psychiatric disorders. This project is ideal if you like programming, data mining, analyzing cutting-edge genomics data, and interpreting the biological results. You'll work closely with computational and genomic scientists, medical doctors, attend group meetings, and work in a high-collaborative interdisciplinary team. We hope you can join us!
Deep Learning for Imaging-Genomics-Transcriptomics Single-cell Integration
"Computational Biology Projects"
Faculty Advisor: Manolis Kellis
Mentor(s):
Contact e-mail: kellis-admin@mit.edu
Research Area(s): Computational Biology
The MIT CompBio lab (compbio.mit.edu) headed by Prof. Manolis Kellis is
looking for talented SuperUROP students for computational projects in
single-cell genomics, epigenomics, and disease circuitry. Expertise in
programming, data analysis, method development, and machine learning will
be great strengths for a successful project.

We have recently generated some of the largest collections of single-cell
datasets of human disease, across Alzheimer's, Schizophrenia, Bipolar, Down
Syndrome, Heart Disease, Obesity, Exercise, and other brain, metabolic, and
cardiac tissues and disorders. We seek computational SuperUROP students to help
analyze these datasets in close collaboration with researchers in our team
in the context of a truly interdisciplinary team, where highly-talented
computational postdocs and students work side-by-side with each other, and
with leading experimental scientists to design the next generation of
statistical techniques, computational tools, experimental datasets,
integrative models, and validation experiments. You'll be working with
state-of-the-art tools for analyzing single-cell genomics and epigenomics
datasets, developing new methods for their integration, and working with
other scientists in the group to understand their biological implications.

To get an overview of our recent results in these areas, here's a recent
talk by Professor Kellis: Link and the corresponding slides: Link

For more background information please visit Link
Computational Biology Projects
"Learning Stochastic Chemical Kinetics Models of Eukaryotic Transcription"
Faculty Advisor: Alan Edelman
Mentor(s): Alexander LeNail, Chris Rackauckas
Contact e-mail: lenail@mit.edu
Research Area(s): Computational Biology, Inference, Machine Learning
Transcription is regulated by sequence-specific DNA-binding proteins called Transcription Factors, which cooperatively configure chromatin loci to be either recalcitrant to or amenable to RNA polymerization. To actuate endogenous transcriptional programs at will, we need a model for how these Transcription Factors (TFs) configure chromatin and facilitate transcription across the genome. This project will develop such a model, and the inference algorithms to do so.

In particular, this project will (1) develop parameter fitting tools in the Julia programming language for Stochastic Chemical Kinetics, (2) fit the parameters of detailed ODE models of transcription to high-throughput single cell ATAC-seq and RNA-seq data and (3) compare the learned parameters of the ODE model with experiments of single cell RNA dynamics in the context of TF perturbations.
Learning Stochastic Chemical Kinetics Models of Eukaryotic Transcription
"Is it time to retire the ring road?"
Faculty Advisor: Professor Cathy Wu
Mentor(s):
Contact e-mail: cathywu@mit.edu
Research Area(s):
The ring road represents a simple model of a highway, which replicates important traffic phenomena such as stop-and-go waves. In recent years, the model has been instrumental in advancing our understanding of the effectiveness of autonomous vehicles (AVs) for congestion mitigation. At the same time, there is a huge gap between this simple model and the real world. There simply aren't any ring roads in real traffic systems. With advances in AI, computing, and mobility data collection, we can now revisit how close (or far) this simple model is from reality. This project seeks to get to the bottom of this question, by examining the consistency of the simple highway model as compared to models with varying degrees of realism. Is it time to retire the ring road?
"A mechanistic model for roadway safety"
Faculty Advisor: Professor Cathy Wu
Mentor(s):
Contact e-mail: cathywu@mit.edu
Research Area(s):
Something curious happened during COVID-19. Driving decreased, but accidents and fatalities **increased**. Driving without seatbelts is an incomplete explanation because the phenomenon occurred far more substantially in the US than many other countries. An alternative hypothesis is that, without traffic congestion to slow things down, the wide roadways in the US resulted in high speeds and thus more crashes. This project is about digging deeper into this phenomenon and constructing a mechanistic model to explain it. We consider COVID-19 as a natural experiment and leverage accident data to investigate what happens when people drive the way the system is designed. A mechanistic model for roadway safety
"Safety as a mobility performance measure"
Faculty Advisor: Professor Cathy Wu
Mentor(s):
Contact e-mail: cathywu@mit.edu
Research Area(s):
Autonomous vehicles (AVs) are anticipated to bring forth immense benefits in terms of safety and performance, such as congestion, emissions, and travel time consistency. However, few methodologies are available to consider the two goals together and their possible trade-offs. As AV technology matures, such more nuanced analysis is required to inform remaining AV development and adoption, including business and policy decisions. To reconcile fundamental differences between safety and performance measures, this project is about devising a statistical measure of safety that allows safety to be analyzed alongside performance in microscopic traffic models. The new notion of safety in turn induces a family of optimal control problems corresponding to a multitude of accident-prone traffic microsimulation scenarios, in which AVs can be optimized according to different risk tolerance levels. Safety as a mobility performance measure
"TranViT: An Integrated Vision Transformer Framework for Real-Time Transit Travel Time Prediction"
Faculty Advisor: Professor Jinhua Zhao
Mentor(s): Dr. Awad Abdelhalim
Contact e-mail: jinhua@mit.edu
Research Area(s):
The aim of this project is to efficiently integrate and utilize image data from roadside cameras, transit vehicle feeds, and other urban mobility data sources to improve the accuracy of transit travel duration and arrival time prediction. This project can be sub-divided into two main research tasks:

- Developing a comprehensive framework for the integration of roadside traffic images and transit feeds.

- Tasks include image acquisition, manipulation, and labeling to be used for vision model training.

- Labels are generated based on the observed traffic state from transit vehicles' feed processed from transit travel time.

- We will investigate different approaches of estimating traffic state labels based on image data, including classification, detection, and a combination of both methods.

- Utilizing traffic state predictions to improve the accuracy of transit travel time prediction

- After training and validating vision model(s) to produce reliable state estimates from images, the state estimates will be used in addition to existing data sources to develop predictive models for transit travel time.

- A comprehensive literature review for the SoTA models in travel time prediction will be conducted. Different statistical, machine learning, and time series models are to be developed and evaluated.
TranViT: An Integrated Vision Transformer Framework for Real-Time Transit Travel Time Prediction
"Robots that can learn on-the-fly"
Faculty Advisor: Vivienne Sze & Sertac Karaman
Mentor(s): Soumya Sudhakar
Contact e-mail: soumyas@mit.edu
Research Area(s): Machine Learning, Robotics
For energy-constrained robots navigating using deep neural networks (DNNs), it is beneficial for the robot to be able to decide when to spend energy on online training to improve the DNN in order to navigate the environment better. The robot can learn "on-the-fly" by using active learning (deciding what subset of inputs to train on) and self-supervised training (training from unlabeled data). The goal of this SuperUROP project is to contribute to the development of algorithms for uncertainty-based active learning and online self-supervised training. The SuperUROP will be involved in the algorithm design and testing of different active learning strategies, as well as the implementation of a self-supervised training framework.

Low-Energy Autonomy and Navigation Group website: Link

Prior experience in DNN training is a plus.
Robots that can learn on-the-fly
"Engineering super-textiles for heat and moisture management in extreme environments"
Faculty Advisor: Svetlana V. Boriskina
Mentor(s): Vlad Korolovych
Contact e-mail: sborisk@mit.edu
Research Area(s):
Managing heat and moisture transport is important in everyday textiles but becomes crucial in applications such as wound care and in extreme environments such as inside a warfighter helmet or a spacesuit. In particular, spacesuit and helmet systems are optimized for protection and enhanced communications with little consideration for comfort and the need for cooling. The risk of overheating can be uncomfortable, disorienting, and deadly. On the other hand, wound care also requires a dressing to quickly absorb and manage moisture and temperature to improve the healing process. The goal of this project is to engineer and test multi-layer multi-purpose textiles that can control transport and capture of heat and moisture. This is a great opportunity to contribute to ongoing research projects for the US Navy and industry and to advance fundamental understanding of heat and mass transport properties on the micro- and nano-scale. Engineering super-textiles for heat and moisture management in extreme environments
"Toward transiently probing near-field heat and momentum transfer"
Faculty Advisor: Svetlana V. Boriskina
Mentor(s): Simo Pajovic
Contact e-mail: sborisk@mit.edu
Research Area(s):
In radiative transfer, the near-field is defined as the regime where the sizes of objects and the distances between them are less than the peak wavelength of thermal radiation, given by Wien's law. The breakdown of the laws of thermal radiation in the near-field regime is well-established, but the transition to the contact regime.that is, from radiation to conduction.remains mysterious. Experimentally, one of the main obstacles to probing this transition is "snap-in," the phenomenon in which two surfaces in close proximity will suddenly snap into contact because of surface forces. The goal of this project is to establish the time scale of this phenomenon and understand whether it is possible to transiently probe radiative heat and momentum transfer during snap-in. The student will be responsible for building on an existing finite difference method (FDM) model of snap-in, exploring refinements to the model such as the proximity force approximation (PFA), and parametric studies to understand the limits of snap-in. This project may lead to experiments that validate the model and provide insight into the transition from radiation to conduction for the first time ever. Toward transiently probing near-field heat and momentum transfer
"Pangenome-wide transcript quantification and differential gene expression analysis in bacteria"
Faculty Advisor: Caroline Uhler
Mentor(s): Ashlee Earl
Contact e-mail: aearl@broadinstitute.org
Research Area(s): Computational Biology, Inference, Machine Learning
RNA-seq is a commonly used high-throughput technique to quantify gene expression. Typically, sequencing reads are mapped to a reference genome to determine their likely gene of origin. Most bacterial species, however, exhibit high variability in gene content, and selecting a single reference that comprehensively represents gene content across all strains is impossible, limiting our ability to accurately quantify and compare gene expression across strains. In this project, we aim to develop new algorithms and statistical methods that can quantify gene expression across a species pangenome, i.e., the set of all gene families observed across a collection of strains. Successful completion of this project will vastly improve differential gene expression analyses, enabling less biased and more complete understanding of the differences in gene expression across a wide variety of species and strains compared to the state-of-the-art. Pangenome-wide transcript quantification and differential gene expression analysis in bacteria
"Generative modeling for fair and explainable AI in bioimage analysis research"
Faculty Advisor: Caroline Uhler
Mentor(s): Juan Caicedo
Contact e-mail: jcaicedo@broadinstitute.org
Research Area(s): Computer Graphics and Vision, Inference, Machine Learning
Many automated decisions in biotechnology centers will soon depend on ML and AI models. However, modern architectures for image analysis are often black boxes that cannot be interpreted in a direct way. This project aims to harness the power of generative models to provide meaningful interpretations of visual differences between biological images. Explainability models will be used to communicate important visual patterns learned by deep neural networks to differentiate between groups of images. We expect these models to enhance the reliability and understanding of image-based automated decision systems in biological research. Generative modeling for fair and explainable AI in bioimage analysis research
"Profiling cell interactions in tissue images using machine vision"
Faculty Advisor: Caroline Uhler
Mentor(s): Juan Caicedo
Contact e-mail: jcaicedo@broadinstitute.org
Research Area(s): Computer Graphics and Vision, Inference, Machine Learning
The next frontier in basic biological research is to crack open the mysteries of how cells work together to form higher level living structures, such as tissues and organs. In this project, we aim to use representation learning algorithms to model the spatial relationships of cells in high resolution tissue images. In particular, we will use vision transformer architectures together with self-supervised learning algorithms to extract patch-level features from tissues at multiple scales. In addition, this project aims to explore optimal transport strategies to efficiently estimate feature matches across two different tissue samples and to identify where they have similar structure and where they differ. Profiling cell interactions in tissue images using machine vision
"Discovering the rules of cellular state transitions in cancer"
Faculty Advisor: Caroline Uhler
Mentor(s): Nir Hacohen
Contact e-mail: nhacohen@broadinstitute.org
Research Area(s): Computational Biology, Inference, Machine Learning
A limitation in cancer therapeutics is the development of therapy resistance. Of particular interest is the ability of cancer cells to exhibit plasticity, or the ability to interconvert from therapy-sensitive to therapy-resistant states under selective pressure. By leveraging state-of-the-art lineage tracing tools, we have shown direct evidence of plasticity in several cancers. However, the rules governing how cells interconvert between states and the trajectories they follow are poorly understood. In this project, students will leverage single-cell multiomic data (transcriptomic and epigenomic) from clonally barcoded cells over a time-course to uncover how cancer cells exhibit plasticity to evade therapy. In particular, students will be mentored by postdocs in the Hacohen lab and computational scientists in the Broad Data Science Platform to develop methods to uncover regulators of plasticity using optimal transport, machine learning and probabilistic models. Discovering the rules of cellular state transitions in cancer
"Discovering the spatial organization and interaction of cells in cancer using high-dimensional data."
Faculty Advisor: Caroline Uhler
Mentor(s): Nir Hacohen
Contact e-mail: nhacohen@broadinstitute.org
Research Area(s): Computational Biology, Inference, Machine Learning
Cancer remains a leading cause of death. While therapies targeting immune system activation have shown promise in certain cancer types, they remain limited in others. We have recently discovered that cancers with favorable response to immune therapies are spatially organized in cellular networks ('hubs') (Pelka, Cell, 2021); however, the rules governing the formation and functions of these hubs are poorly understood and may unlock potential new therapeutic approaches. Students will work with a team of computational, experimental and clinical scientists in the Hacohen lab to dissect how cells are organized and interact with each other in human cancer samples using single-cell transcriptomic and single-molecule resolution spatial transcriptomic datasets. Students will be mentored in developing statistical and machine learning methods for integration of multimodal data and extracting features underlying the rules of cellular spatial organization. Discovering the spatial organization and interaction of cells in cancer using high-dimensional data.
"Evaluation of neural networks for object segmentation in biological images"
Faculty Advisor: Caroline Uhler
Mentor(s): Beth Cimini
Contact e-mail: bcimini@broadinstitute.org
Research Area(s): Computer Graphics and Vision, Inference, Machine Learning
While deep neural networks (DNNs) have made amazing strides in segmentation of photographic/natural images over the last several years, progress has been slower in scientific images. Recent tools such as Cellpose/Omnipose and StarDist have made strides in the segmentation of nuclei and some kinds of cells, but retraining still requires large amounts of data, and existing interfaces are difficult to use for many computationally uncomfortable users. Our new tool, Piximi, is designed to run DNNs from a friendly web interface. Major open questions still exist, however, as to what networks and training strategies work best in this space, especially for cases where ground truth is very limited. In this project we will explore various biological use cases, network types, hyperparameters, and preprocessing options to create default models that will allow scientists to easily segment their images. Evaluation of neural networks for object segmentation in biological images
"Generative sequence models anticipate pathogen immune markers"
Faculty Advisor: Caroline Uhler
Mentor(s): Martin Strazer, Daniel Graham
Contact e-mail: mstrazar@broadinstitute.org,dgraham@broadinstitute.org
Research Area(s): Computational Biology, Inference, Machine Learning
Pathogens and the human adaptive immune system are in constant arms race driven by evolution. Here, failure to recognize true pathogen markers, or recognizing self through molecular mimicry may lead to catastrophic health outcomes. Human T cells recognize non-self cells by binding to their small molecules, proteins or peptide sequences, collectively termed antigens. Existing peptide antigen prediction models rely on sequence classification, which interpolate within existing experimental data. In contrast, generative sequence models promise to anticipate previously unseen sequence markers, which resembles how humans retain a constantly evolving pool of T cell receptors. The project goal is to design generative sequence models, or adversarial architectures, trained on hundreds of thousands of peptides recognized by human immune cells. Established proteomics and synthetic biology protocols will provide training data and validation of computational predictions. Antigen prediction models are critical to anticipate novel, self-similar microbial peptides leading to autoimmune disease, inform vaccines targeting cancer or emerging pathogens. Generative sequence models anticipate pathogen immune markers
"Probabilistic annotation of microbial metabolites through mass spectrometry"
Faculty Advisor: Caroline Uhler
Mentor(s): Martin Strazer, Ramnik Xavier
Contact e-mail: mstrazar@broadinstitute.org,rxavier@broadinstitute.org
Research Area(s): Computational Biology, Inference, Machine Learning
Human gut microbes possess a tremendous repertoire of yet unexplored metabolic capabilities. Discovery of novel bacterial metabolites, enzymes and mechanisms is hindered by sparse annotation of molecules surveyed through mass spectrometry. The project will provide an opportunity to design probabilistic models, graph neural networks or learning of molecular representations to better identify known and predict novel molecular structures. Modeling of dependencies between molecules in a spectrum may be augmented further using laws of chemistry, databases of reactions and pathways, as well as sequenced microbial genomes. Direct experimental data and feedback will be provided by testing the predictions using an in-house collection of hundreds of bacterial isolates. The proposed models aspire to reduce the turnaround time in discovery of microbial chemistry, which in turn informs development of therapeutics and probiotics relevant to bowel inflammation and cardiovascular disease. Probabilistic annotation of microbial metabolites through mass spectrometry
"earning Brownian motion trees and applications to cell differentiation"
Faculty Advisor: Caroline Uhler
Mentor(s):
Contact e-mail: cuhler@mit.edu
Research Area(s): Computational Biology, Inference, Machine Learning
Every cell in our body contains the same genetic information. However, we have many different cell types and they all show different gene expression patterns. In order to understand how this is possible, we will study how the different gene expression patterns develop, starting from a stem cell and differentiating into the different cell types. We will assume that gene expression develops along the tree according to a multivariate Brownian motion with correlation among the genes. Given gene expression data from various points along the unknown differentiation tree, the goal of this project is to develop algorithms that can simultaneously map the cells to the differentiation tree and learn the underlying tree topology. What statistical guarantees can be obtained? How well does the algorithm perform on simulations? Does it provide meaningful differentiation trees when applied to real biological data? earning Brownian motion trees and applications to cell differentiation
"Causal inference and reinforcement learning"
Faculty Advisor: Caroline Uhler
Mentor(s):
Contact e-mail: cuhler@mit.edu
Research Area(s): Computational Biology, Inference, Machine Learning
In many applications the end goal of causal inference is not necessarily to learn the underlying causal system but to infer the best interventions in order to push the underlying system towards a desired state. This is the case for example when studying reprogramming, where the goal is to determine the best interventions (e.g. over-expression of particular transcription factors) to push a differentiated cell towards the stem cell state. In this project, the goal is to build on methods in active learning, RL and causal inference to obtain methods for selecting the best interventions in order to push the system towards a desired state. Causal inference and reinforcement learning
"Learning causal graphs and applications to gene regulation"
Faculty Advisor: Caroline Uhler
Mentor(s):
Contact e-mail: cuhler@mit.edu
Research Area(s): Computational Biology, Inference, Machine Learning
Causal inference is a cornerstone of scientific discovery because it asks why? Most methods for learning causal graphs assume that the underlying graph is a DAG, i.e., that it does not contain any directed cycles. However, feedback loops in biological networks are not only common but also crucial features. In addition, many causal inference algorithms do not allow for imposing prior knowledge on the directed graph or cannot be applied to large networks. In this project, the goal is to develop causal inference algorithms that can overcome these limitations and can be applied to infer gene regulatory networks. These networks have about 20'000 nodes, but there is a lot of prior information on the network coming from knock-out experiments. In order to increase the power of the methodology, it is important to be able to use this prior information. What statistical guarantees can be obtained and what is the computational trade-off? How well does the algorithm perform on simulations? Does it provide meaningful gene regulatory networks when applied to real biological data? Learning causal graphs and applications to gene regulation
"Freeform contact visualization based on updated Archimedes' principle"
Faculty Advisor: Nicholas Fang
Mentor(s): Zhiguang Liu
Contact e-mail: liuzg@mit.edu, nicfang@mit.edu
Research Area(s):
Greek scientist Archimedes taught us that, a totally or partially immersed object suffers from a buoyant force equal to the weight of the fluid displaced by itself. This well-known principle offers an ingenious way for weight measuring and supports related inventions, such as ship and submarine. In recent years, researches of water-walking arthropods, which expel water by hydrophobic legs instead of immersing into it, update the boundary of Archimedes' principle, say floating force that holds the arthropods equals to the expelled volume of water due to surface tension, but it is extremely challenging to perform real-time measurement and analysis for included mechanical status because of the micro-newton or lower weight order. To solve this problem, former studies employed the sunk water surface to bend the light and create extended shadow which can well visualize the tiny force. However, previous works are mainly in the scope of point or uniform 1D contact with water, the research of freeform hydrophobic contact which is closer to realistic condition is currently lacking. We aim to explore the relationship between critical parameters of free-floating thread and its visualized counterpart via both experimental and simulated approaches. The goal is to develop an effective model to quantitatively describe the contact by resolving the inverse problem.
Image source: Link
Resource: 18.357 Interfacial Phenomena Link
Freeform contact visualization based on updated Archimedes' principle
"Enhancing Flexibility and Resilience of the Future Grid with Coordinated IoT Integration"
Faculty Advisor: Anuradha Annaswamy
Mentor(s): psrivast@mit.edu
Contact e-mail: aanna@mit.edu
Research Area(s):
Environmental and sustainability concerns are resulting in a rapid penetration of distributed energy resources into the power grid. Coupled with the recent advancements in Internet-of-Things (IoT) technology, this is steering the electric grid away from the traditional unidirectional pathways towards an interconnected web with edge computing capabilities. Recent cyber attacks on the power grid have been of increasing complexity and sophistication, and ensuring the resiliency of the grid to service critical infrastructure is of utmost importance. On one hand, with the integration of IoTs into the system, attacks of various degrees are imminent. On the other hand, with proper and secure coordination structures to leverage the ubiquitous presence of IoT nodes in place, the flexibility and resiliency of the power grid could be enhanced. The UROP project will focus on the development of a hierarchical electricity market with coordinated and meaningful aggregation of IoT enabled assets at various voltage levels to provide system operators with overall awareness about the grid operating conditions. The proposed framework will be validated on a realistic power system model to provide routine grid services under normal operating conditions and also demonstrate the increase in critical load served in the face of contingencies. Enhancing Flexibility and Resilience of the Future Grid with Coordinated IoT Integration
"Retail market models in a low-carbon electricity grid"
Faculty Advisor: Anuradha Annaswamy
Mentor(s): rhaider@Mit.edu
Contact e-mail: aanna@mit.edu
Research Area(s):
With increasing pressures to decarbonize the electricity grid, the grid edge is witnessing rapid adoption of customer-owned devices (rooftop solar, EVs). When coordinated appropriately, these devices can help increase grid efficiency and operational flexibility. Programs like net energy metering (NEM) pay resource owners for generated electricity; however these payment structures are inefficient and too expensive. Companies which aggregate the capabilities of small-scale resources are also becoming increasingly prevalent. These aggregators participate directly in wholesale electricity markets (WEM) and make decisions on behalf of resource owners. Such participation poses problems as device setpoints determined by aggregators may violate local grid constraints, and the business model of companies may reduce net profits for resource owners. This project concerns the development of a retail electricity market wherein resources participate directly, at the local level. Prior work has shown that such a market is financially more attractive than NEM and more importantly accounts for local grid constraints unlike aggregators. The project will consist of an in-depth comparison of the aggregator model and retail market. Development of models for retail-WEM interactions and aggregators, and use of these components in a case study with real and synthetic data are some of the key project objectives. Retail market models in a low-carbon electricity grid
"Feedback Control for Robot Manipulation"
Faculty Advisor: Russ Tedrake
Mentor(s): none
Contact e-mail: russt@mit.edu
Research Area(s): Control and Decision Systems
Despite it's prevalence in almost every other field of systems engineering, shockingly almost no state of the art systems for robot manipulation actually use feedback control (because we lack the basic theory). The MIT Robot Locomotion group is pursuing fundamental research in this direction, and applying it to physical robots.

The project will involve (convex/mixed-integer) optimization, some systems theory, and lots of robots -- mostly programmed in python. We don't expect you to know all of that coming in, but hope you're excited about them.

The project may also involve robot perception systems -- developing feedback directly from pixels is a major intellectual theme of our current work.
"Fleet learning for manipulation"
Faculty Advisor: Russ Tedrake
Mentor(s): none
Contact e-mail: russt@mit.edu
Research Area(s): Control and Decision Systems
It takes a long time for one robot to generate enough data to learn a new manipulation skill. But that's not actually the problem we need to solve -- soon we will have many similar robots performing some distribution of tasks in a distribution of environments. Might this fundamentally change the way that we program our robots today?

We have a new project that aims to investigate the rigorous theoretical underpinnings of fleet learning for robotics, including distributed learning, multi-task learning, and learning from non-iid data. Depending on the interest/aptitude of a SuperUROP candidate, there could be room to contribute to the theory, the software implementation, and or to hardware robot demonstrations.
"Manipulation of Soft Objects"
Faculty Advisor: Russ Tedrake
Mentor(s): none
Contact e-mail: russt@mit.edu
Research Area(s): Control and Decision Systems
The MIT Robot Locomotion Group has been making progress in connecting machine learning perception with robotics planning and control to enable an increasingly diverse set of manipulation capabilities for robotics. (e.g. Link )

We are now working to extend this work to deal with increasingly complex systems -- using machine learning perception, potentially learning intuitive physics, and rigorous dynamics and control + lots of robots.
"Precision engineering for the world's most precise force sensor"
Faculty Advisor: Vivishek Sudhir
Mentor(s): Dominika Durovcikova
Contact e-mail: vivishek@mit.edu
Research Area(s):
Precision force sensing has reached zeptonewton scales (1 zN =10^(-21) N) in recent years thanks to a combination of advances in quantum measurements and precision engineering. We are building an experiment to push the frontiers of force sensing even further - to the sub-yoctonewton (<10^(-24) N) regime - by using a single trapped electron as a force transducer. Force detection is enabled through continuous measurement of the electron's displacement by coupling its motion to a microwave cavity field via image currents.

The realization of this force sensor requires integration of the electromagnetic trap operating in the GHz regime with a microwave resonator, and coupling the two via an antenna, while maintaining precision alignment of the trap components with low levels of added noise. The super-UROP project will focus on the precision design and assembly of the mechanical structure.
Precision engineering for the world's most precise force sensor
"Process Simulatior for GaN Devices"
Faculty Advisor: Tomas Palacios
Mentor(s):
Contact e-mail: tpalacios@mit.edu
Research Area(s): Materials, Devices and Photonics, Nanotechnology, Numerical Methods
We are working to develop a custom semiconductor microfabrication process simulator that is capable of exploring novel fabrication techniques. This software allows us to accurately model fabrication flows and test the results with existing device physics simulators. The current iteration of our simulator is designed using python, but a student with additional knowledge of C programming would help improve efficiency to enable more accurate simulation routines. Additionally, we would like to add a complete user interface to the software to transform the scripting based program into an intuitive and user friendly application. Through this SuperUROP position, a student will sharpen their programming skills while learning about the physics of all major processing steps in semiconductor fabrication. Process Simulatior for GaN Devices
"Mechanical Circulatory Support Hemodynamic Effect on the Aorta and Clotting"
Faculty Advisor: Elazer R. Edelman
Mentor(s): Angela Lai, Farhan Khodaee
Contact e-mail: angelai@mit.edu, farhank@mit.edu
Research Area(s):
Veno-arterial extracorporeal membrane oxygenation (VA-ECMO) is used to support patients with cardiopulmonary failure. A cannula returns oxygenated blood to the vasculature in the opposite direction of cardiac output. This project will study the effects in the mixing zone, where antegrade pulsatile and continuous retrograde flow meet in the aortic truck, on the recovery of aortic endothelial cells and their subsequent effect on coagulation. Weaning off of ECMO is non-standard across hospitals and we will aim to scale and model this clinical phenomenon in a microfluidic chip lined with human aortic endothelial cells. Participation in this project will bring the student a strong foundation in experimental design, tissue engineering, and an opportunity to learn more about the blood-material interface of medical devices. Ultimately, we hope to graduate this project to a small animal model. Students with a background in biology, mechanical engineering, tissue engineering, and related disciplines are encouraged to apply.

Research Area(s): Tissue Engineering, Medical Device Optimization, Hemodynamics
Mechanical Circulatory Support Hemodynamic Effect on the Aorta and Clotting
"AI-driven inverse design of lens systems"
Faculty Advisor: Faez Ahmed
Mentor(s):
Contact e-mail: faez@mit.edu
Research Area(s):
Deep generative models, such as PaDGAN, are proven to be a useful tool for creating novel high-performance designs. This projects aims to expand their application to a new domain of optical lens design, which has applications from phone cameras to space applications. The goal is to enable machine learning algorithms to instantly generate design of new lens systems that meet multiple engineering requirements.

The student will work on training new deep generative models on lens datasets and explore how to incorporate manufacturing and optical constraints into the methods. You will be working with a group of researchers from MIT and EPFL. Image source: Link
AI-driven inverse design of lens systems
"Training models of the COIN Hardware"
Faculty Advisor: Cardinal Warde
Mentor(s): none
Contact e-mail: warde@mit.edu
Research Area(s): Computer Architecture, Computer Networks, Machine Learning, Signals and Systems, Theoretical Computer Science
The Photonic Systems Group is developing a low-power Compact Opto-electronic Integrated Neural (COIN) network co-processor for applications such as image recognition and classification, and sensor fusion. The COIN, which is inspired by biology, employs arrays of photodetectors and low-power thresholding circuits in silicon, in combination with arrays of light emitters and holographic optical interconnection elements sandwiched together into a brick. We are exploring the development of such a neural network system first at the discrete component level first, before committing to the chip-based integrated version.

This project involves the development and implementation of neural network algorithms (including weight perturbation) to train software models of the COIN, prior to uploading the weights to the actual COIN hardware. Also, studies attempting to quantify the fault tolerance properties of these networks are ongoing.
Training models of the COIN Hardware
"Assembly and Testing of Hybrid COIN Network"
Faculty Advisor: Cardinal Warde
Mentor(s): none
Contact e-mail: warde@mit.edu
Research Area(s): Applied Physics, Circuits, Computer Architecture, Computer Systems, Materials, Devices and Photonics, Signals and Systems
The Photonic Systems Group is developing a low-power Compact Opto-electronic Integrated Neural (COIN) network co-processor for applications such as image recognition and classification, and sensor fusion. The COIN, which is inspired by biology, employs arrays of photodetectors and low-power thresholding circuits in silicon, in combination with arrays of light emitters and holographic optical interconnection elements sandwiched together into a brick. We are exploring the development of such a neural network system first at the discrete component level first, before committing to the chip-based integrated version.

This project involves: (1) design and modeling of the overall discrete-component neural network system (contains photodetectors, LEDs, MOSFETS, holographic optical elements and micro-controllers), and (2) assembly and testing of the system with real-world input after upload of the weights. Follow-on work would focus on contributing to the design and fabrication of the final compact VLSI integrated-circuit version of the COIN.
Assembly and Testing of Hybrid COIN Network
"2-D Bistable Optical Array"
Faculty Advisor: Cardinal Warde
Mentor(s): none
Contact e-mail: warde@mit.edu
Research Area(s): Applied Physics, Circuits, Computer Architecture, Materials, Devices and Photonics
The Photonic Systems Group is developing a low-power Compact Opto-electronic Integrated Neural (COIN) network co-processor for applications such as image recognition and classification, and sensor fusion. The COIN, which is inspired by biology, employs arrays of photodetectors and low-power thresholding circuits in silicon, in combination with arrays of light emitters and holographic optical interconnection elements sandwiched together into a brick. We are exploring the development of such a neural network system first at the discrete component level first, before committing to the chip-based integrated version.

This project involves: (1) a search for exotic materials that inherently exhibit low-power optical bistability, (2) design, modeling and assembly of existing low-power opto-electronic components to form an associated optical bistable subsystem, and (3) testing of the resulting hybrid optical bistable array. Follow-on work would focus on contributing to the design and fabrication of the final compact VLSI integrated-circuit version of the COIN.
2-D Bistable Optical Array
"More than one model for learning: a computational investigaation"
Faculty Advisor: Professor Robert C. Berwick
Mentor(s):
Contact e-mail: berwick@csail.mit.edu
Research Area(s): Natural Language and Speech Processing
We make decisions in every moment of our lives. Among the very first, and perhaps the most important one, is the language we learn. Every child is capable of learning every language in the world. It is now clear that all languages share many structural regularities.a Universal Grammar.and slo clear that the specific language(s) children learn can only be fixed by input in their linguistic environment.

This project investigates the mechanism of making linguistic decisions: How do children integrate the general constraints on language with the language-specific input? A prominent proposal likens children to "little scientists" who weigh data.the specific linguistic input-against competing scientific hypotheses, selecting the optimal one(s). In a recent paper, the learner is endowed with primitive computational processes akin to string operations. Learning consists of evaluating the combinations of these primitives against the input, a set of examples drawn from a formal language such as a regular languages or context-free languages. The combination(s) with the highest posterior probabilities are selected as winners. A number of important questions, both technical and linguistic, remain; these will be explored in this SuperUrop aiming to find out what computational model best fits what children actually do.
More than one model for learning: a computational investigaation
"To be determined based on student interest"
Faculty Advisor: Professor Tal Cohen
Mentor(s):
Contact e-mail: talco@mit.edu
Research Area(s):
Professor Cohen is looking to supervise a SuperUROP and will work with a student one-on-one to determine a project based on the student's interest. For more information about the Cohen Mechanics Group, please check out this article in The Tech: Link and the group website: Link To be determined based on student interest
"Biomateriomics and Deep Learning for Sustainable Materials"
Faculty Advisor: Professor Markus Buehler
Mentor(s):
Contact e-mail: mbuehler@mit.edu
Research Area(s):
In this project, students will use a host of multiscale modeling tools to study, discover and manufacture sustainable materials using a range of techniques including atomistic simulations, deep learning, generative design, as well as manufacturing and material characterization. Through computational engineering, materials can be analyzed and built across scales, from nano to macro, and manufactured in the lab. These types of new materiomic approaches are revolutionizing the way we source and recycle materials through molecular re-engineering, for applications ranging from structural engineering to food and agriculture. Specific areas of interest include biological and bio-inspired materials, as well as naturally sourced living organisms like mushroom mycelia, farmed and engineered for de novo architected materials designs. Students will work in a multidisciplinary lab with cutting edge computational and experimental facilities. The project can be tailored depending on the student's specific interests. Biomateriomics and Deep Learning for Sustainable Materials
"Scaling and Accelerating Hardware Simulation"
Faculty Advisor: Daniel Sanchez
Mentor(s): Fares Elsabbagh
Contact e-mail: farese@csail.mit.edu, sanchez@csail.mit.edu
Research Area(s): Circuits, Computer Architecture, Computer Systems, Systems (incl OS, databases, computer security)
Current chips built with modern technology are incredibly complex, packing tens of billions of transistors organized in hundreds to thousands of blocks. These chips have become so large that we can't simulate their behavior quickly enough. Slow simulation increases hardware design time and limits testing, making hardware bugs more likely.

This situation is paradoxical---after all, as transistor budgets have grown, processors have also become more capable, so in an ideal world simulation speed should have kept pace with chip complexity. However, software simulation speed has barely improved in the past 15 years, because it is hard to parallelize: simulation requires fine-grained parallelism, dividing work into small tasks with few operations each that communicate frequently. This is a poor match to existing multicores, which work well only with coarse-grained tasks that communicate infrequently.

This project seeks to accelerate hardware simulation (specifically, RTL simulation) by building on Swarm, an architecture designed to accelerate hard-to-parallelize applications that provides hardware support for fine-grain parallelism. We have already ported the fastest simulator to this system and achieved speedups of over 100x. But much remains to be done, including new software optimizations to avoid needless work (e.g., simulating only events that change signal on a given cycle) and exploit more parallelism, and hardware specialization techniques to improve simulation performance and efficiency. We are building a large multi-FPGA prototype of this system with over 1000 cores, and expect speedups of over 1000x, running a month-long simulation in ten minutes.
Scaling and Accelerating Hardware Simulation
"An Architecture to Accelerate Computation on Encrypted Data"
Faculty Advisor: Daniel Sanchez
Mentor(s): Nikola Samardzic
Contact e-mail: nsamar@mit.edu, sanchez@csail.mit.edu
Research Area(s): Computer Architecture, Computer Systems, Systems (incl OS, databases, computer security)
Fully Homomorphic Encryption (FHE) is a class of encryption schemes that allow computing on encrypted data. FHE enables secure computation offloading. For example, a client can use FHE to send encrypted deep learning inference requests to a server, which the server can execute while keeping all data encrypted. FHE guarantees privacy even if the server is untrusted or compromised. Unfortunately, FHE carries substantial performance overheads: it is about 10000x-100000x slower than computing on unencrypted data. This has so far limited the adoption of FHE.

In this project, you will work with a team of students to design a specialized accelerator for FHE computations that achieves speedups of over 10000x over software, bridging most or all of FHE's performance gap, as well as new compiler and software techniques to use this new hardware well. In preliminary work (F1, MICRO 2021, and CraterLake, ISCA 2022) we have designed accelerators that achieve speedups of over 5000x over CPUs and are fast enough to run large deep neural networks in milliseconds instead of minutes. There are many opportunities to improve performance further through better hardware and software techniques. Given the broad scope of this project, there are many areas where you can contribute depending on your interests, including FPGA and ASIC prototyping, compiler techniques, and hardware-algorithm co-design.
An Architecture to Accelerate Computation on Encrypted Data
"A Hardware Accelerator for Sparse Computations"
Faculty Advisor: Daniel Sanchez
Mentor(s): Yifan Yang, Axel Feldmann
Contact e-mail: yifany@csail.mit.edu, axelf@csail.mit.edu, sanchez@csail.mit.edu
Research Area(s): Computer Architecture, Computer Systems
Conventional processors, like CPUs and GPUs, have been designed and optimized for regular applications that operate on structured data, like dense matrices. However, many emerging application domains, such as deep learning, graph analytics, and sparse linear algebra, do not fit this model: these applications perform irregular and unstructured operations on large, sparse data structures, like sparse matrices or tensors. Existing processors cannot execute these applications efficiently, as they incur excessive data movement and cannot use regular processing elements like vector units efficiently.

In this project, you will work with a team of students to design a hardware accelerator for these irregular algorithms. This accelerator uses a new hardware/software interface based on sparse tensors that provides an efficient, unified abstraction for applications in these seemingly disparate domains (e.g., allowing us to build a single accelerator that runs graph and deep learning applications efficiently). Moreover, this accelerator features computational units that make irregular algorithms memory-friendly by restructuring their computation. This approach spends hardware on cheap compute operations to reduce expensive data movement, eliminating the memory traffic bottleneck of conventional architectures.
A Hardware Accelerator for Sparse Computations
"Cracking the code of pattern selection in drying suspensions"
Faculty Advisor: Irmgard Bischofberger
Mentor(s): Paul Lilin
Contact e-mail: irmgard@mit.edu,plilin@mit.edu
Research Area(s):
Have you ever wondered why paint cracks and peels from your wall? Why mud fractures during dry periods? And why the crack patterns look similar in both of these seemingly very different systems? Paint and mud are two examples of suspensions of colloidal particles, small particles suspended in a solvent. The intricate drying patterns that occur in such suspensions, which are seen in deserts around the world, have also been used as microfabrication templates and as tools to identify diseases from a single dried drop of blood serum. However, a fundamental understanding of the crack length-scale selection and of the crack branching process is currently lacking. This experimental project aims to identify the laws that relate the thickness profile of the dried deposit to the crack pattern using a model colloidal suspension, with potential applications in both forward-problems applicable to manufacturing (what drying conditions achieve a desired crack pattern?) and backward-problems applicable to disease detection (what are the suspension properties that created a specific crack pattern?). By participating in this project, you will learn state-of-the-art confocal imaging and manufacturing techniques and have opportunities to develop skills in fracture mechanics and poroelasticity.
Research Area(s): pattern formation, fluid dynamics, www.mitfluidslab.com
Cracking the code of pattern selection in drying suspensions
"Physics-based learning for ultrafast laser sources"
Faculty Advisor: Sixian You
Mentor(s):
Contact e-mail: sixian@mit.edu
Research Area(s): Applied Physics, BioEECS, Machine Learning, Materials, Devices and Photonics, Signals and Systems
This project aims to develop physics-based learning algorithms for optimizing ultrafast laser sources. Ultrafast laser sources have deep impacts in imaging, sensing, and quantum computing. However, very few existing laser sources can adaptively tune to the task with new spatiotemporal profiles. The student will join forces with the postdoc and graduate students in the lab to develop this new algorithm-enabled laser source. Physics-based learning for ultrafast laser sources
"Holography for neural stimulation and AR/VR"
Faculty Advisor: Sixian You
Mentor(s):
Contact e-mail: sixian@mit.edu
Research Area(s): Applied Physics, BioEECS, Energy, Power, Electromagnetics, Machine Learning, Materials, Devices and Photonics, Signals and Systems
This project aims to find new ways to generate faithful and high-speed 3D holograms using optics and algorithms. Holograms have wide-reaching impacts on AR/VR, 3D printing, optical manipulation, lithography, and neural stimulation. However, it remains an unsolved problem how to generate faithful 3D holograms using 2D light shaping devices. This project will explore new ways of designing the algorithms and/or the optics to improve the 3D hologram generation for neural stimulation and near-eye displays. Holography for neural stimulation and AR/VR
"Optically Gated GaN Power Transistors"
Faculty Advisor: Tomas Palacios
Mentor(s): none
Contact e-mail: tpalacios@mit.edu
Research Area(s): Materials, Devices and Photonics, Nanotechnology
GaN power transistors have been widely adopted in power electronics systems, offering lower parasitics, therefore higher switching speed and efficiency. This project proposes an optically gated GaN transistor which seeks to overcome the limitations of conventional GaN power transistors (electrically gated transistors). In the proposed device, electrical isolation is achieved between the gate and the channel. Furthermore, the switching speed would be determined by the generation of electron hole pairs (and possibly also rate of avalanche), potentially allowing for significantly higher switching frequency. More significantly, the proposed concept would represent a step forward in electronics-photonics integration on III-N platform for optical interconnects, photonics interposer etc. The eventual goal is to study the feasibility of integrating the proposed device in gate driver circuits. Understanding of operation of microelectronic device and opto-electronic device is desired.
"Gallium Nitride CMOS Technology for the Next Generation of RF Front End Modules"
Faculty Advisor: Tomas Palacios
Mentor(s):
Contact e-mail: tpalacios@mit.edu
Research Area(s): Materials, Devices and Photonics, Nanotechnology
The realization of a high-data-rate uplink/downlink for 5G NR and beyond relies on the development of mm-wave transmitters, whose performance is in turn dominated by the linearity and power efficiency of the power amplifiers (PAs). Even though gallium nitride (GaN) PAs have delivered unprecedented levels of power at Ka band and beyond, their limited linearity force the transistors to operate at high power back-off, which significantly reduces the overall amplifier efficiency. Furthermore, to support the high data rates for 5G, the bandwidth of the system needs to be widened. New bottom-up approaches, based on innovations at the semiconductor technology-level, are highly desired to fundamentally improve linearity and efficiency in PAs for the next generation of RF front end modules.

To this end, the use of GaN complementary type (CMOS) technology, an emerging development at the semiconductor technology-level, is proposed. GaN CMOS technology pioneered at MIT has achieved leading performance and is a promising candidate for future monolithically integrated applications, but more could be done at the device-level to improve the performance. At the circuit-level, proof-of-concept simulations reveal significant improvement in the linearity metrics for a PA using GaN CMOS. Further simulation studies will be conducted to evaluate the device-circuit interaction and design trade-offs in these systems. This project also includes plans to prototype the circuit designs based on micro-fabrication at MIT.nano. Desired areas of knowledge include: operation of transistor, analog/RF circuit, micro-fabrication.
Gallium Nitride CMOS Technology for the Next Generation of RF Front End Modules
"Gallium Nitride RF Electronics for Application in W-Band Telecommunication and Quantum Computing Systems"
Faculty Advisor: Tomas Palacios
Mentor(s):
Contact e-mail: tpalacios@mit.edu
Research Area(s): Materials, Devices and Photonics, Nanotechnology
Gallium nitride high electron mobility transistors (HEMTs) has been the workhorse of 5G telecommunication systems thanks to its ability to deliver unprecedented levels of power at mm-wave frequencies. This excellent performance is attributed to the superior combination of properties of GaN, including wide band gap (3.4 eV), high critical electric field (5 MV/cm2), high electron mobility (2000 cm2/V&#61655;s) and the design of the III-N heterostructure. Therefore, it is not surprising that, GaN HEMTs are extensively used in the transmitters in 5G base stations.

Future telecommunication systems beyond 5G FR2 would require operation at higher frequencies and wider bandwidths to transmit the enormous amounts of data. To this end, fundamental improvements to today's GaN HEMTs are required to allow these transistors to sufficient power at higher frequency (W band and beyond) with higher RF linearity. Several innovations at the device-level have been proposed, including epitaxial structure and device design. Furthermore, the use of GaN electronics has been proposed for application in quantum computing systems, where the electronic components interface the qubits. The SuperUROP student will be exposed to the design, micro-fabrication (at MIT.nano) and RF characterization of the device prototypes. Knowledge of transistor operation and micro-fabrication is desired.
Gallium Nitride RF Electronics for Application in W-Band Telecommunication and Quantum Computing Systems
"ML-guided intubation"
Faculty Advisor: Thomas Heldt
Mentor(s):
Contact e-mail: thomas@mit.edu
Research Area(s): BioEECS, Computer Graphics and Vision, Machine Learning, Signals and Systems
Intubation, especially when required to do emergently, is a high-stakes medical procedure that requires visualization of the oropharynx and larynx to place a breathing tube into the patient's trachea. The advent of video laryngoscopy allowed for the development of sizeable data archives of successful and failed intubation approaches. This project seeks to leverage such a data archive to develop computer-vision algorithms to aid physicians in the timely placement of breathing tubes by automatically detecting anatomical landmarks in real-time and suggesting trajectories for optimal placement of the endotracheal tube. This project involves a collaboration with clinical colleagues from Boston Children's Hospital Department of Emergency Medicine. ML-guided intubation
"Acoustic-cue-based speech analysis for recognition"
Faculty Advisor: Stefanie Shattuck-Hufnagel
Mentor(s): Jeung-Yoon Elizabeth Choi
Contact e-mail: sshuf@mit.edu
Research Area(s): Cognitive AI, Communications, Human Computer Interaction, Machine Learning, Natural Language and Speech Processing, Signals and Systems
This SuperUROP project aims to develop a speech analysis system based on a theory of human speech perception, building on detection/interpretation of individual acoustic cues to words and sounds. Its has 2 components: system development, and analysis of acoustic patterns. Applications include extension to different languages/dialects, atypical speech, and prosody (pitch/timing patterns).
Preparation: A required summer training program, (40 hours over a period of 2-4 weeks, remote participation possible) on critical aspects of speech acoustics, feature-cue labelling, signal processing and analysis, preparing for project start in Fall 2022. NOTE: Candidates who have undergone this training previously are exempt from this requirement.
Opportunity 1: Speech analysis system developer. Code modules for automatic recognition and parsing of acoustic cues to distinctive features of phonemes and words. Requires some experience with Python or similar; Matlab, signal processing and/or machine learning will be a plus. Opportunity to learn about speech signal processing, modeling techniques and cues to speech sounds.
Opportunity 2: Speech modification profiler. Analyze modifications in acoustic cues across different speech contexts. Requires experience with Python; general alignment techniques will be a plus, as will background in linguistic phonetics. Opportunity to discover principles that govern systematic modification patterns in communicative speech.
Acoustic-cue-based speech analysis for recognition
"Detecting and Subtyping Cleft Palate"
Faculty Advisor: Marzyeh Ghassemi
Mentor(s):
Contact e-mail: mghassem@mit.edu
Research Area(s): Machine Learning
Spoken language processing can serve as an easy-to-apply diagnostic tool and can be used for screening for clinical trials, surgery risk assessment etc. [1] Two major aspects of speech can be measured by extracting different features from speech and it's associated transcript: what is being said ("content"), and how it is being said ("articulation").
We are particularly interested in the "articulation" component for detecting impaired speech. Specifically, detect impairment in the spectrum of speech produced by children with cleft lip and palate. Cleft lip and cleft palate are openings or splits in the upper lip, the roof of the mouth or both. Prevalence is high among children [2], and speech is affected in distinct ways: hypernasality, weakened pressure consonants, etc. [1]
Speech sounds, just like any other sound, are rapid fluctuations in air pressure. The properties of the produced speech waves are of great importance since they constitute the basis for both the phonetician's acoustic analysis and auditory transcription [3]. Since phonetics form an important component of cleft palate diagnosis/treatment, characteristics of voice are important indicators, which we can extract from speech audio.
In collaboration with the Boston Children's Hospital (BCH), our goal is to build a machine-learning system to accomplish three tasks: distinguish cleft palate speech from healthy speech, classify it into subtypes, and predict Velopharyngeal insufficiency (VPI) scores (annotated by clinicians) all using speech only.
The main project components involved would be:
- Develop speech-based audio segmentation and processing pipeline
- Develop phone-level pronunciation scoring pipeline
- Develop machine learning models to classify speech signals
- Run data collection studies to collect normative examples of healthy speech
Detecting and Subtyping Cleft Palate
"Addressing racial and gender biases in the lung transplant allocation system"
Faculty Advisor: Marzyeh Ghassemi
Mentor(s):
Contact e-mail: mghassem@mit.edu
Research Area(s): Machine Learning
There are a large number of well-established biases in the health system that adversely affect the quality of care received by minority populations. While the increasing use of algorithms and machine learning (ML) in healthcare holds great promise, it also risks codifying and worsening these disparities. In this project, we study a process that is particularly prone to racial and other inequities: the organ transplant allocation system. Specifically, we focus on lung transplantation, aiming both to characterize disparities in the current system and suggest ML-based approaches to improve the fairness of outcomes and access.
The UROP will use statistical analysis to make inferences about allocation fairness, build machine learning models to predict physician decisions on transplant acceptance, and run counterfactual simulations to gain further insights. Experience with statistical inference and machine learning, especially in R or Python, is strongly preferred.
Addressing racial and gender biases in the lung transplant allocation system
"Investigating Rotation Prediction"
Faculty Advisor: Pulkit Agrawal
Mentor(s):
Contact e-mail: pulkitag@mit.edu
Research Area(s): Robotics
Predicting rotations is a generically useful skill across many robotic tasks. However, it turns out predicting object rotation from pointcloud data is very challenging. Some reasons for the task difficulty are: multi-modality induced by object symmetry and discontinuities in rotation representations such as quaternions.

In this project, we will conduct a thorough investigation of various rotation representations to design a better methodology for predicting rotations. Such an algorithm is likely to have widespread use as a plug-and-play library across robotics.

Prerequisites:

- Background in RL and ML.
- Mathematical and Programming Maturity.
- Familiarity with Tensorflow or PyTorch.

Look at our recent projects here: (Link
Investigating Rotation Prediction
"Commanding Robots with Natural Language"
Faculty Advisor: Pulkit Agrawal
Mentor(s):
Contact e-mail: pulkitag@mit.edu
Research Area(s): Control and Decision Systems, Natural Language and Speech Processing, Robotics
Imagine opening a website that streams a video of a robot in action and a language prompt that takes in instructions. Building a robotic system that can act from natural language instructions has been a long-standing challenge that will push a step closer towards the dream of a house robot.

Recent developments in reinforcement learning, representation learning, and large scale language models have opened up several avenues for making substantial progress.

We have several ongoing projects in the lab around this theme, and we are happy to discuss more specific ideas with interested students. The projects range from infrastructure development, self-supervised robot learning, language models, connecting language with robotic skills etc.

Prerequisites:
- Background in RL, ML and/or Robotics.
- Mathematical and Programming Maturity.
- Familiarity with Tensorflow or PyTorch.

Look at our recent projects here: Link
Commanding Robots with Natural Language
"Exploring Generative Models for Multimodal Control"
Faculty Advisor: Pulkit Agrawal
Mentor(s):
Contact e-mail: pulkitag@mit.edu
Research Area(s): Machine Learning, Robotics
Generative models have been producing amazingly crisp images or uncannily real paragraphs of text. However, these models haven't made their way into robot control. In multimodal policy learning, we care about learning a model that is able to ingest (state, action) tuples where a single state may have many corresponding optimal actions. For example, when a human demonstrates a task to a robot, it's common for humans to perform the task in different ways. However, this creates problems for machine learning models trying to mimic the human. In this project, we will develop robotic systems that can learn from diverse and multi-modal human demonstrations.

Prerequisites:
- Background in RL and ML.
- Mathematical and Programming Maturity.
- Familiaity with Tensorflow or PyTorch.

Look at our recent projects here: Link
Exploring Generative Models for Multimodal Control
"Reinforcement Learning for Robot Dancing"
Faculty Advisor: Pulkit Agrawal
Mentor(s):
Contact e-mail: gmargo@mit.edu
Research Area(s): Robotics
Some robotic tasks naturally define suitable rewards, for example touching a point with a robot arm or walking in a target direction for a legged robot. However, other tasks do not admit such clean reward definitions. One such task is dancing, where a sequence of distinct motions should be executed in an aesthetically pleasing style. Boston Dynamics is beloved for its videos of Spot dancing, and it would be fun to see if such behaviors can be obtained by simply specifying high-level rewards instead of manually hard-coding different moves.

Look at our recent projects here: Link

Prerequisites:
- Background in RL and ML.
- Mathematical and Programming Maturity.
- Familiaity with Tensorflow or PyTorch.
Reinforcement Learning for Robot Dancing
"Artificial Glial Neural Network for Reinforcement Learning"
Faculty Advisor: Pulkit Agrawal
Mentor(s):
Contact e-mail: pulkitag@mit.edu
Research Area(s): Cognitive AI, Control and Decision Systems
Deep Neural Networks were inspired by network of neurons in the brain. For a long time, Glia cells were thought to perform only metabolic function such as cleansing or supplying energy to neurons. However, new research has shed light that glia cells are critical for long-term memory and exploration.

In collaboration with neuroscientists, in this project we will attempt to bridge recent findings about Glia-Cells with Deep Learning to create a new family of artificial neural networks, dubbed Glia-Deep Neural Networks (Glia-DNNs).

Glia DNNs are expected to be substantially better at long-term reasoning, multi-task learning and avoiding catastrophic forgetting.

Prerequisites:
- Background in RL and ML.
- Mathematical and Programming Maturity.
- Familiaity with Tensorflow or PyTorch.
Artificial Glial Neural Network for Reinforcement Learning
"Investigating Human Preferences via Multi-Sensory Perception"
Faculty Advisor: Pulkit Agrawal
Mentor(s): none
Contact e-mail: pulkitag@mit.edu
Research Area(s): Cognitive AI
Information from various sensory modalities is tightly coupled with each other. The sight of strawberries evokes its sweet taste, listening to vroom-vroom sound instantly conjures images of a motorbike or a car! There is plenty of evidence that humans represent the world by integrating information from multiple sensory modalities. This project aims to tease about the exact nature of these representations and explore how predictable are touch, audio, vision and taste signals are from each other. For e.g., is it the case that people who prefer rock music also prefer beer over wine?
"Self-Supervised Cloth Manipulation"
Faculty Advisor: Pulkit Agrawal
Mentor(s): none
Contact e-mail: pulkitag@mit.edu
Research Area(s):
Manipulating deformable objects such as clothes is a significant challenge in robotics. This project will investigate how a robot can autonomously collect data to understand various ways in which it can change the configuration of clothes. These models will be then be used to fold clothes and perform other manipulation tasks. We will make use of and advance deep learning, model-based reinforcement learning and imitation learning methods. The most relevant prior work is the lab’s previous project on rope manipulation: Link
"Improving Credit Assignment with sequence models"
Faculty Advisor: Pulkit Agrawal
Mentor(s):
Contact e-mail: pulkitag@mit.edu
Research Area(s): Control and Decision Systems, Machine Learning, Robotics
A substantial challenge in reinforcement learning (RL) is credit assignment. Because an agent may only get a reward after it performs a sequence of actions, it may not be possible to identify which action was most important. Such inability to identify important actions (i.e., assign credit) leads to many problems in RL such as data inefficiency, training instability etc.

In this project, we will explore the use of sequence models such as Transformers, S4 model to improve credit assignment.

Expected Results: If successful, the project will have a widespread impact in improving data efficiency across the board of state-of-the-algorithms. Essentially, better decisions with less data.

Prerequisites: Background knowledge in reinforcement learning and machine learning, mathematical and programming maturity, and is comfortable using deep-learning frameworks such as PyTorch or Tensorflow.

Look at our recent projects here: Link
Improving Credit Assignment with sequence models
"Molecular Devices: Fabrication and Design"
Faculty Advisor: Farnaz Niroui
Mentor(s): none
Contact e-mail: fniroui@mit.edu
Research Area(s): Applied Physics, Materials, Devices and Photonics, Nanotechnology
Molecules with precisely defined intrinsic properties are attractive nanoscale building blocks for the development of novel devices with functionalities that would not be feasible using conventional silicon-based technologies. With dimensions less than 5 nm though, integration of molecules into active devices is a great challenge and has hindered progress in the field. In this project, we will develop techniques that allow us to fabricate arrays of molecular structures with sub-nanometer precision and resolution over large areas. These techniques will provide a platform based on which we can design various molecular devices with emerging applications in electronics, optics and sensing.
"Reinforcement Learning for Predicting and Shaping Electricity Demand from Electric Vehicles"
Faculty Advisor: Mardavij Roozbehani
Mentor(s):
Contact e-mail: mardavij@mit.edu
Research Area(s): Control and Decision Systems, Energy, Power, Electromagnetics, Machine Learning, Signals and Systems
Unmanaged demand from a highly concentrated number of Electric Vehicles (EVs) can cause demand spikes and stress the grid. As we move towards adoption of Electric Vehicles (EVs) at larger scale, managing the electricity demand from these vehicles becomes more necessary. In this project we are interested in developing reinforcement learning models and algorithms that help us estimate the aggregate demand from a fleet of EVs from limited information. This information is typically in the form of past consumption patterns and potentially responses to incentives and prices at an aggregate (fleet) level. We also investigate how much additional information at individual vehicle level (for instance, arrival time, charging deadline, desired charging level) can improve accuracy of estimation, thereby quantifying the value granular of data. We will then build on the predictive model to develop pricing and incentive algorithms to shape the aggregate response from a fleet of EVs to match supply and reduce stress on the grid. Strong background in ML, DNN, and reinforcement learning is preferred. Reinforcement Learning for Predicting and Shaping Electricity Demand from Electric Vehicles
"Reinforcement Learning for Agriculture"
Faculty Advisor: Munther Dahleh
Mentor(s):
Contact e-mail: dahleh@mit.edu
Research Area(s): Control and Decision Systems, Machine Learning
Developing countries particularly in sub-Saharan Africa face a large gap in productivity and agricultural yield compared with western countries. While absence of advanced technology and machinery contribute to this gap, suboptimal agro-management practices are also a factor that need to be addressed. We propose to use real and synthetic data to train a Deep Q-Network to learn automated policies for optimal decision making in irrigation and fertilization. Such decisions can be then tailored and specialized to a specific farm and crop based on the known parameters. This project builds on our previous work in which we built a reinforcement learning model that could learn how to provide daily agro-management recommendations based solely in information from the weather, training on simulated crop yields and policies that were generated using the WOFOST model. In this phase we aim at developing a model that can incorporate sensory feedback from the levels of nutrients and moisture in the soil. This is a particularly challenging learning problem because the reward (yield) evolves with a much slower dynamics compared to the rest of the system parameters. In addition, for practical purposes we are interested in developing simplified strategies that require infrequent and/or quantized actions. Reinforcement Learning for Agriculture
"Embedding Invisible Markers into Textiles"
Faculty Advisor: Prof. Stefanie Mueller
Mentor(s): Doga Dogan
Contact e-mail: stefanie.mueller@mit.edu
Research Area(s): Computer Graphics and Vision, Human Computer Interaction, Materials, Devices and Photonics
This project will explore how we can embed unobtrusive markers into textiles/clothes for tracking human posture and motion. Markers such as ArUco or OptiTrack reflective markers are currently used for human motion capture (e.g., for movies and animation generation). However, they are visible and are not useable in practical applications where aesthetics are important. This project will propose new ways to hide these markers for different interactive applications. This will involve the use of infrared-absorbing inks and infrared-reflective materials.

Useful skills and experience: 3D printing, spraying, manufacturing, OpenCV or other image processing libraries, Python or similar coding experience. Optional: Machine learning and CNN

Related projects:
- InfraredTags: Link
- DefeXtiles: Link
Embedding Invisible Markers into Textiles
"Dynamic QR Codes"
Faculty Advisor: Prof. Stefanie Mueller
Mentor(s): Doga Dogan
Contact e-mail: stefanie.mueller@mit.edu
Research Area(s): Computer Graphics and Vision, Human Computer Interaction, Materials, Devices and Photonics
This project will investigate the design and fabrication of printed QR codes that can change their content (i.e., bits) without using any battery and electronic components. The student will investigate the use of thermochromic, photochromic, and/or electrochromic inks to create printed code bits that change color based on a physical change in the environment or the attached object (e.g., temperature or brightness change).

Useful skills and experience: Digital fabrication/manufacturing methods, materials, spraying, electronics, rapid prototyping, OpenCV or other image processing libraries, Python or similar coding experience.

Related projects:
- InfraredTags: Link
- DefeXtiles: Link
Dynamic QR Codes
"Computational design tools to build greater material usage awareness"
Faculty Advisor: Prof. Stefanie Mueller
Mentor(s): Mackenzie Leake
Contact e-mail: stefanie.mueller@mit.edu
Research Area(s): Computer Graphics and Vision, Human Computer Interaction
Digital tools can support the design and prototyping of a wide range of physical objects from clothing to city buildings. When going through the design process, at different stages designers need to be more or less aware of the visual appearance of the object and its material and physical constraints, and it can be quite challenging to balance these competing considerations. This project will focus on two areas: 1) designing tools that help users become more aware of material usage at key moments in the design and construction process and 2) developing algorithms for helping users create lower waste designs. Applications for these software tools might include sewn garments and furniture, but we would also be happy to discuss other domains based on your creative interests.

Helpful skills and interests include:
Web development
Computational geometry
Optimization
Computational fabrication (e.g., CAD, laser cutting)
Any type of creative domain expertise (or curiosity) (e.g., sewing, painting, woodworking, etc.)
Computational design tools to build greater material usage awareness
"Stability/Safety of Deep Neural Networks for Dynamical Systems"
Faculty Advisor: Luca Daniel
Mentor(s): Wang Zang
Contact e-mail: luca@mit.edu
Research Area(s): Control and Decision Systems, Energy, Power, Electromagnetics, Machine Learning
Deep neural networks (DNNs) have become widely adopted in a variety of machine learning tasks, including modeling and control of physical systems in a reinforcement learning (RL) setting. For example, in guided policy search, a neural network policy is used to replace the online predictive control policy and directly control a physical system in a feedback loop. In modeling the physics, deep neural networks show great potential in generalizing high dimensional, complex systems such as fluid dynamics. Supported by high-quality database, the statistical way of modeling decision making shows the potential to outperform classical methods. Yet, recent studies demonstrate that neural networks can be surprisingly vulnerable to adversarial attacks and even to non-adversarial random input noise. This suggests that current DNNs may not be as robust and trust-worthy as users hope or expect. Indeed, there could be many potential problems when such networks are employed in the real world without proper care. In particular, for applications that are safety-critical, such as natural gas processing and distribution industry, the existence of instability in neural networks has raised severe and unprecedented concerns - a small perturbation on the neural network input may cause significant change in the control actions, leading to potential destabilization of the system, or driving the system state to an unsafe region. In our group we are working in general toward the development of tools that can verify if some properties related to safety and robustness are satisfied for a given dynamical system, as well as . toward enhancing the safety and robustness of DNNs in the feedback-loop if found to be lacking such crucial properties.
This specific superurop project would involve training a neural network model based on process data provided by a collaborating natural gas company, that captures the dynamical behavior of part of their processing and distribution plants. Furthermore, in a second stage of the project we would analyze the stability properties of the trained model and apply our techniques under development to correct any unreliable and potentially unsafe regions of operation.
Stability/Safety of Deep Neural Networks for Dynamical Systems
"Acoustic Surfaces for Virtual Spaces and Enhanced Environments"
Faculty Advisor: Luca Daniel, Jeff Lang, Vladimir Bulovic
Mentor(s): Jinchi Han, Jose Serralles, Richard Swartwout
Contact e-mail: luca@mit.edu
Research Area(s): Circuits, Environment, Human Computer Interaction, Nanotechnology
This superurop project would be part of a much larger ongoing effort aimed at the development of a nearly imperceptible acoustic technology, designed to enhance the comfort, health, and aesthetics of living spaces. Our design is a floor-to-ceiling wallpaper-like acoustic sensors/speakers that can provide a truly immersive acoustic experience, for life-like sound generation, extraneous noise cancellation, and a healthier acoustic space. Such an umbrella project includes several aspects from the development of microfabricated devices, to the electrical powering and actuation of such devices, all the way to the development of high level algorithms that control them to achieve the desired acoustic experiences.
The superurop project would be more limited in scope. More specifically this superurop project would focus on helping us develop some practical demos, using traditional speakers controlled by some already developed initial algorithms. The student background and interests required for this part of the project would include working experience with electrical wiring and testing of electronic devices, digital to analog interfaces, potentially also FPGAs, filters, amplifiers etc. If the initial demos are promising, the project could continue in a second stage with the circuit development for electronic powering and control of the actual microfabricated devices (as opposed to the initial traditional speakers), or alternatively with the development of more sophisticated algorithms (potentially even including the use of deep neural networks).
Acoustic Surfaces for Virtual Spaces and Enhanced Environments
"Toolkit for the design of reflective tools for learning maker skills"
Faculty Advisor: Prof. Stefanie Mueller
Mentor(s): Dishita Turakhia
Contact e-mail: stefanie.mueller@mit.edu
Research Area(s): Human Computer Interaction
In this project, we will build an interface that helps designers with building and prototyping tools for learning of maker skills using the idea of reflection-in-action. For example, imagine a "talking screwdriver" that helps the users reflect on their performance as they construct a physical object, such as a table or a chair. Reflection in action is an approach that is known to help with learning of concepts and improving skillsets. In this project, we will be focusing on the skills related to making and fabrication, such as lasercutting, 3D printing, soldering, using power tools, etc. We will prototype several tools for enabling learning of these maker skills through reflection-in-action. This will involve using appropriate sensors, deploying algorithms for prompting reflection in the users, and assessing learning gains in the users. We will also develop an interface (web tool using HTML Javascript or stand-alone tool using Python) to enable designers prototype the design of such reflection-in-action tools. In order to work on this project, the skills required will be - prototyping, electronics, programming. Toolkit for the design of reflective tools for learning maker skills
"Hydration Monitoring Device"
Faculty Advisor: Luca Daniel and Martha Gray
Mentor(s): none
Contact e-mail: luca@mit.edu
Research Area(s): Applied Physics, BioEECS, Signals and Systems
Our hydration project is focused on developing the technology missing for reliable and physiologically meaningful hydration tracking, with an aim of optimising hydration and avoiding dehydration. We created the project at MIT with a primary focus on the substantial need for improved hydration management in the elderly, but with a view to the needs in many sectors and the broader population. We are currently carrying out clinical testing in Boston of non-invasive technologies in a prototype. Examples of prjects that are just starting up include:
1) Designing a portable set up that would allow for at home testing of the hydration sensor;
2) Explore possible model systems that could be used to support sensor calibration; understand the use of IPA, and possibly provide an alternative as a means of calibrating the device;
3) Validating hardware set up, including: stability, temperature dependence, motion sensitivity. Work may evolve to include signal processing approaches.
"Analysis of EMG Signal and Motion Tracking Data in Running"
Faculty Advisor: Luca Daniel & Brian Anthony
Mentor(s): Praneeth Namburi
Contact e-mail: luca@mit.edu
Research Area(s): BioEECS, Cognitive AI, Control and Decision Systems, Human Computer Interaction
Biological movements, such as breathing as walking, are incredibly efficient. Efficient storage and retrieval of elastic energy by the elastic tissues of the body (muscles, tendons, ligaments, and other connective tissue) is a key reason for efficient movement. This knowledge is used intuitively by athletes, martial artists, and dancers. The goal of this project is to understand this process scientifically, which can have a wide impact, including improvement of athletic training and physical therapy programs.
Several simple models have been used by scientists in hopes of understanding the movements and techniques used by the professionals. Recent developments in motion capture and 3D modeling technology allow us to test more complex models of muscle-bone systems and further our understanding of complex movements.
In this project you will use the space and tools of the MIT.nano Immersion Lab to analyze how the signal and energy travels through muscle systems during running - one of the essential movements of the human body.
"Efficient Algorithm Design for Training Deep Neural Networks"
Faculty Advisor: Luca Daniel
Mentor(s): Lam M. Nguyen
Contact e-mail: luca@mit.edu
Research Area(s): Machine Learning, Numerical Methods, Theoretical Computer Science
Training of some deep neural networks can be challenging since the loss surface of their network architecture is generally non-convex, or even non-smooth. Indeed, there has been a long-standing question on how optimization algorithms used in training can be made to converge to a global minimum. In general, existing optimization algorithms cannot guarantee the convergence to a global solution without convexity or other strong assumptions. In this project we could explore several potential questions along this line of research. For instance:
1) Designing practical algorithms for training neural networks that may converge to global solutions.
2) Studying optimization frameworks for machine learning tasks with a theoretical guarantee to global optimal solutions given some reasonable assumptions. This could be done for instance by taking advantage of some special structure of neural network architectures.
Additional references can be provided to any student that is potentially interested in these research directions and would like to discuss further. We are also open to tailoring the final research plan to the student's specific background and interests, provided that the student's existing background and interests include strong theoretical and mathematical aspects.
Efficient Algorithm Design for Training Deep Neural Networks
"Robot Swarms for 3D Printing"
Faculty Advisor: Prof. Stefanie Mueller
Mentor(s): Martin Nisser
Contact e-mail: stefanie.mueller@mit.edu
Research Area(s): Human Computer Interaction, Machine Learning, Robotics
This project will explore how to integrate swarm robots into a closed-loop fabrication pipeline. Our lab has a swarm of cubic two-wheeled robots capable of precise localization, and have a ramp with which they can locomote onto the build plate of a 3D Printer. The goal of this project is to develop algorithms and/or hardware to enable novel applications using this feature, and builds on our recently developed rapid fabrication process for robots ("Hermits" and "Laserfactory", links below).

Applications and goals for this project may include 1) 3D Printing new mechanical interfaces which swarm robots can mate with to create on-demand, heterogeneous robots. 2) Integrate swarms into the 3D printing process by utilizing them as "dynamic" support material, removing existing supports, or assembling 3D printed parts on the platform. 3) fabricating shells for on-demand, heterogeneous swarms.

Useful skills/experience include Manufacturing (3D printing, CAD, electronics) and Robotics (dynamics, SLAM, control theory) and Machine learning, particularly reinforcement learning.

Links below:

Hermits: Link

LaserFactory: Link

Phd Supervisor: Link
Robot Swarms for 3D Printing
"Acoustic-cue-based speech analysis for recognition"
Faculty Advisor: Stefanie Shattuck-Hufnagel
Mentor(s): Jeung-Yoon Elizabeth Choi
Contact e-mail: sshuf@mit.edu
Research Area(s): Cognitive AI, Communications, Human Computer Interaction, Machine Learning, Natural Language and Speech Processing, Signals and Systems
This SuperUROP project aims to develop a speech analysis system based on a theory of human speech perception, building on detection/interpretation of individual acoustic cues to words and sounds. Its has 2 components: system development, and analysis of acoustic patterns. Applications include extension to different languages/dialects, atypical speech, and prosody (pitch/timing patterns).
Preparation: A required summer training program, (40 hours over a period of 2-4 weeks, remote participation possible) on critical aspects of speech acoustics, feature-cue labelling, signal processing and analysis, preparing for project start in Fall 2022. NOTE: Candidates who have undergone this training previously are exempt from this requirement.
Opportunity 1: Speech analysis system developer. Code modules for automatic recognition and parsing of acoustic cues to distinctive features of phonemes and words. Requires some experience with Python or similar; Matlab, signal processing and/or machine learning will be a plus. Opportunity to learn about speech signal processing, modeling techniques and cues to speech sounds.
Opportunity 2: Speech modification profiler. Analyze modifications in acoustic cues across different speech contexts. Requires experience with Python; general alignment techniques will be a plus, as will background in linguistic phonetics. Opportunity to discover principles that govern systematic modification patterns in communicative speech.
Acoustic-cue-based speech analysis for recognition
"Optimization and analysis of morphing hydrofoils through surrogate modeling"
Faculty Advisor: Wim M. van Rees
Mentor(s):
Contact e-mail: wvanrees@mit.edu
Research Area(s):
Analyzing the hydrodynamic performance and stability of a structural hydrofoil design typically requires the numerical solution of a complex, computationally expensive fluid-structure interaction problem. This approach becomes intractable against current developments in smart, programmable materials and additive manufacturing techniques, which drastically increase the design space and open novel opportunities for passively and actively morphing wings. To fully exploit these capabilities, a new paradigm for analyzing and optimizing aeroelastic structures in high-dimensional parameter spaces is required. This project aims to establish an efficient numerical design approach for elastically morphing structures in aerodynamic flows, relying on training surrogate models to learn the interaction between fluid loading and the shape. The proposed approach is applied to a 2D articulated foil through a combination of inviscid and viscous flow simulations, both steady and unsteady. This project offers exposure to surrogate models like Gaussian process regression, as well as numerical simulation methods for fluid flows and structural mechanics. Optimization and analysis of morphing hydrofoils through surrogate modeling
"Flapping fins with chordwise and spanwise curvature variations for underwater propulsion and maneuvering"
Faculty Advisor: Wim M. van Rees
Mentor(s):
Contact e-mail: wvanrees@mit.edu
Research Area(s):
We aim to investigate the straight-line and maneuvering performance of flapping fins with chordwise and spanwise curvature variations using 2D and 3D numerical simulations. Biologically-inspired flapping fin propulsion has potential for underwater vehicles due to the associated high propulsive efficiency at a range of speeds, agility for operation in confined and debris-filled areas, and a concealed profile. In nature, swimmers achieve these benefits by relying on active, musculature-driven shape deformations of the entire body as well as individual fins. We have recently investigated the effect of fin curvature variations on straight-line swimming, but for maneuvering the optimal kinematics and the associated efficiency remain largely unclear. This project proposes to use numerical simulations with our in-house Navier-Stokes solvers to explore the effect of the relevant parameters, analyze the flow field, and characterize the maneuverability landscape. It offers exposure to unsteady fluid dynamics, numerical simulations, and high-performance computing. Flapping fins with chordwise and spanwise curvature variations for underwater propulsion and maneuvering
"Bluefish: Diagram Construction using Constraints and Gestalt"
Faculty Advisor: Daniel Jackson
Mentor(s): Josh Pollock
Contact e-mail: jopo@mit.edu
Research Area(s): Computer Graphics and Vision, Human Computer Interaction, Programming Languages (incl software eng), Systems (incl OS, databases, computer security)
Diagram Construction using Constraints and Gestalt

People spend a lot of time drawing diagrams, and often can't use automated layout tools because they are too inflexible. We've developed a new approach that lets you express layout rules using Gestalt grouping principles, and then applies an automatic constraint solver to size and position elements.

Qualifications: proficiency in JavaScript; interest in visual layout.

More info: Link
Video: Link
Bluefish: Diagram Construction using Constraints and Gestalt
"Riffle: A New Architecture for Data Intensive Apps"
Faculty Advisor: Daniel Jackson
Mentor(s): Geoffrey Litt
Contact e-mail: glitt@mit.edu
Research Area(s): Human Computer Interaction, Programming Languages (incl software eng), Systems (incl OS, databases, computer security)
We are exploring a radical new architecture that will make it possible to build apps more easily. The key ideas are: (a) "local-first": the app operates on local data, and syncs when online, allowing higher performance and greater flexibility; (b) declarative reactive queries: the user interface is built by associating queries with UI elements which are spontaneously executed when data changes, gaining a simplicity similar to spreadsheet formulas, but the full power of a modern UI; (c) relational storage: all data is stored in a relational database, even the state of the UI (eg, which tabs are open and where a panel is scrolled to). We have built a Spotify clone as an initial prototype, and have shown that basing an app on reactive database queries can be surprisingly fast: SQL queries even execute faster than the browser can update the DOM.

Qualifications: familiarity with a JavaScript webstack (as taught in 6.170) and preferably SQL too.

More info: Link
Riffle: A New Architecture for Data Intensive Apps
"Dark Patterns, Privacy and Policy Conformance"
Faculty Advisor: Daniel Jackson
Mentor(s):
Contact e-mail: dnj@mit.edu
Research Area(s): Human Computer Interaction, Programming Languages (incl software eng), Systems (incl OS, databases, computer security)
Many apps use "dark patterns" to trick users into taking actions they might not have chosen; privacy concerns are rising amongst consumers; and new regulations are appearing (with GDPR, and new bills in Congress and many states). How should software be designed to be fair and honest with users, but still meet a company's needs? We are developing a new approach based on concept design that provides criteria for deciding when designs are "dark" and suggests design moves for meeting the app developer's needs while respecting the user.

In collaboration with Danny Weitzner (Internet Policy Research Initiative).

Qualifications: interest in user-facing design, and preferably 6.170.
Dark Patterns, Privacy and Policy Conformance
"Extremism and Social Media"
Faculty Advisor: Fotini Christia
Mentor(s): Chris Hays and Zach Schutzman
Contact e-mail: cfotini@mit.edu
Research Area(s):
n recent years, revelations about the impact of social media on society have motivated an interesting research direction: What are the contributing factors on social media platforms that drive certain users towards extremism, like eating disorders, white nationalism or religious radicalism? This project explores how algorithmic and network-based forces can contribute to extremism online. To probe the characteristics of social media algorithms and communities within their user base, we will orchestrate a large number of synthetic social media accounts (bots) with features similar to real-world users. Then, we will selectively expose some bots to extremist content to see how the bots and their communities are affected. The first phase of the project will be focused on creating realistic bots that have profiles and demonstrate friending, consuming, engagement and sharing behavior similar to social media platforms' real users.

Possible areas where the student will contribute: (Prior experience with any of these is a plus but not required.)
- Statistical / ML methods to generate the behavior of a synthetic social media user (bot), including training NLP text generation models, simulating realistic location data, and other challenges
- Data collection:
o Tools to collect large amounts of public data about social media users, like location, bio, number of friends, frequency/amount of content consumption/creation per day, etc.
o Tools to collect fine-grained private data about social media users, possibly including an app or browser extension, with their consent (e.g., their friending patterns and page interactions)
- Data processing and misc. other technical challenges related to orchestrating bots at scale
Extremism and Social Media
"Nonlinear Integrated Optics"
Faculty Advisor: Rajeev Ram
Mentor(s): Gavin West
Contact e-mail: rajeev@mit.edu
Research Area(s): Applied Physics, Materials, Devices and Photonics
Nonlinear optics allows you to fuse low-energy infrared photons into visible photons. We have recently demonstrated photonic integrated circuits that are able to generate visible light even with very low power infrared light. The ability to generate new colors of light is a powerful new capability for integrated photonics. The applications include biological sensing and spectroscopy, building atomic clocks, and programming atomic quantum computers.

We were able to achieve these results by using high-voltages to reconfigure the crystal structure of waveguide materials. We are able to configure the structure on microscopic length scales and in this way control the speed with which photons of different colors travel in the waveguide - this enhances the nonlinear interactions and allows us to realize wavelength conversion at eye-safe power.

This project will introduce students to advanced concepts in photonic integration, nonlinear microscopy, tunable laser systems, and solid-state physics. You will learn to control laser systems and convert their light into new colors. This feels like magic.
Nonlinear Integrated Optics
"Chip-scale Microscopes and Mesoscope"
Faculty Advisor: Rajeev Ram
Mentor(s): Marc de Cea
Contact e-mail: rajeev@mit.edu
Research Area(s): Applied Physics, BioEECS, Machine Learning, Materials, Devices and Photonics
The goal of this work is to develop a microscope that is approximately the same size as the object being imaged. The combination of micron-scale LEDs and megapixel CMOS imagers alongside deep prior image processing algorithms allows us to build lens-free microscopes and camera capable of <10 um resolution that can be mm-to-cm in size. Conventional microscopes tend to become larger as the desired spatial resolution increases (for magnification). These large, bench top microscopes with micron-scale resolution typically have limited fields of view - requiring you to scan across the field of view, sequentially capture images, and stitch the image together.

Our architecture fundamentally changes these design trade-offs associated with microscopy. The size of the object being measured dictates the size of the overall microscope that is needed: If you are trying to image a mm-size object with micron-scale resolution, then we can build a mm-scale microscope. If you are trying to image a several-cm tissue sample, we can build a microscope that is also cm-size while preserving the micron-scale resolution. This is all enabled by the unique properties of our nanoLEDs - they have high-spatial coherence and generate nearly spherical wavefronts. The recorded images are digital holograms that we can process computationally to extract 3D features. A new class of physics-based deep learning algorithms works with the low-power digital holograms to realize reconstructed images.

Building these prototypes microscopes/mesoscopes will develop skills in photonic systems (LEDs, holography), 3D printing, and physics-based ML. The biology/medical applications motivating the development of these tools will be used to generate the samples and data sets for measurement and analysis.
Chip-scale Microscopes and Mesoscope
"TinyFL: Federated learning on Edge Devices"
Faculty Advisor: Lalana Kagal
Mentor(s):
Contact e-mail: lkagal@csail.mit.edu
Research Area(s): Computer Systems, Human Computer Interaction, Machine Learning, Systems (incl OS, databases, computer security)
Federated learning (FL) is a technique that allows distributed clients to collaboratively learn a model without sharing any data. It enables clients to train high-quality and robust models without having to share their data. At DIG, we have developed several protocols for FL that are communication efficient, support heterogeneous data, and allow for personalization. In this project, we are focusing on enabling FL across embedded devices with the aim of providing privacy-preserving and communication efficient learning on the edge. The resultant framework could be deployed for sensors in wet labs, home automation and even self driving cars. Some experience with machine learning is required. TinyFL: Federated learning on Edge Devices
"Tricorder"
Faculty Advisor: Rajeev Ram
Mentor(s): Nili Persits
Contact e-mail: rajeev@mit.edu
Research Area(s): Applied Physics, BioEECS, Machine Learning, Materials, Devices and Photonics
We now have the technology to build a tricorder.

A hand-held chemical analyzer that can tell us the concentration and chemical composition of everything from dilithium to drugs, foods, and alien substances. These analyzers employ new photonic tools to perform Raman spectroscopy. The architecture developed in our lab allows us to miniaturize the instruments already used on the Mars 2020 rover into a low-cost, handheld device. This technology can be used for applications as diverse as environmental monitoring, vaccine production, forensic investigation, and surgical guidance. Building these prototypes will develop skills in photonic systems (lasers, fiber optics, and photon counting), sensing circuits and micros, 3D printing, and ML to analyze the huge data flows generated by such a tool.
Tricorder
"Towards Reliable Decision Making"
Faculty Advisor: Aleksander Madry
Mentor(s):
Contact e-mail: madry@mit.edu
Research Area(s): Machine Learning
The goal of this project is to build understanding and tools that help us improve the reliability of our ML models, especially with respect of a variety of distribution shifts. The focus will be on principled modeling and thorough experimental studies.

Note: This project will require very good implementation skills as well as strong ML background and mathematical maturity.
Towards Reliable Decision Making
"Understanding the Impact of the Training-Time Design Choices on ML Model Biases"
Faculty Advisor: Aleksander Madry
Mentor(s): Logan Engstrom, Guillaume Leclerc
Contact e-mail: madry@mit.edu
Research Area(s): Machine Learning
This project aims to build a principled understanding of how different training-time design choices, such as model architecture, learning rate schedule, and set of data augmentations, impact downstream performance of the resulting models. We will want to move here beyond the granularity of just test set accuracy and uncover trends in how different choices change the learned biases at both the individual-example (e.g., one image of a car) and subpopulation (e.g., all red cars) level. Understanding the Impact of the Training-Time Design Choices on ML Model Biases
"Natural Language Descriptions of Deep Networks"
Faculty Advisor: Jacob Andreas & Antonio Torralba
Mentor(s): Evan Hernandez
Contact e-mail: jda@mit.edu
Research Area(s): Machine Learning, Natural Language and Speech Processing
We are developing tools that generate natural language descriptions of learned feature in deep networks. [ Link ]

This UROP project will focus on extending the generality of these tools and their usefulness to human users. Possible directions include:

- extending these techniques to support new model classes (e.g. for natural language processing or robotics).

- studying and improving their usefulness for downstream human decision-making (e.g. automatically surfacing surprising or sensitive features, generating explanations of individual decisions).
Natural Language Descriptions of Deep Networks
"Neural Video Conferencing"
Faculty Advisor: Mohammad Alizadeh
Mentor(s): Vibhaa Sivaraman
Contact e-mail: vibhaa@mit.edu
Research Area(s): Computer Graphics and Vision, Computer Networks, Computer Systems
Video conferencing has become critical to our daily lives, but today's systems continue to suffer from poor quality due to variable and congested network connections. We have been developing a neural video conferencing system that leverages advances in computer vision techniques to produce high-quality videos from low-bitrate encodings. In such a system, when the network fails to provide the desired performance during a session, rather than merely reduce video quality, the clients use pre-delivered neural models to reconstruct a high-quality video of each participant from facial landmark locations, or from even just the audio. Although deep learning approaches have been able to reduce bandwidth requirements compared to traditional codecs, there are two significant areas where the state-of-the-art is lacking. Most prior work has focused on generic video compression and deep learning approaches, rather than exploit the specific characteristics of video calls such as similar backgrounds and repeated participants. Further, deep learning is compute-intensive and running them in real-time on laptops and smartphones is a significant challenge. In this project, we will explore ways to trade-off bandwidth and compute to design a neural video conferencing approach that is feasible on current devices, yet dramatically improves video conferencing quality in bandwidth-constrained environments. Neural Video Conferencing
"Machine Learning for Silicon Photonics Simulation"
Faculty Advisor: Duane Boning
Mentor(s): Zhengqi Gao
Contact e-mail: boning@mtl.mit.edu
Research Area(s): Machine Learning, Materials, Devices and Photonics, Numerical Methods
The Finite-Difference Time-Domain (FDTD) method is a state-of-the-art approach for simulating the behavior of light inside a silicon photonic component. It first divides the simulation region by a large number of grids (also known as Yee cells) and then iteratively updates the electric and magnetic fields according to the discretized version of Maxwell's equations. However, when it comes to the photonic circuit level (i.e., several photonic components connected in sequence), FDTD simulation is almost impossible to perform due to long running time and large memory requirement. Nevertheless, a circuit-level FDTD simulator is of high interest to both academia and industry, to support the design of larger scale integrated silicon photonics. In this project, we attempt to make circuit-level FDTD simulation feasible. Noticing the recent advances in using neural networks to solve partial differential equations (e.g., neural operators), we will use FDTD simulation data of small photonic components to train a neural network, and then apply it to circuit-level FDTD simulation. We will also investigate the generalization ability of the neural network in this process. Through this project, the SuperUROP student will gain a deep understanding of both silicon photonic circuits and machine learning. Machine Learning for Silicon Photonics Simulation
"3D geometry understanding using point cloud sequences"
Faculty Advisor: Justin Solomon
Mentor(s): Xiangru Huang
Contact e-mail: xiangru@mit.edu
Research Area(s): Computer Graphics and Vision, Machine Learning
We will work on tasks in 3D geometry understanding using point cloud sequence data. In contrast to single-frame point clouds, point cloud sequences contain much richer information such as motion, temporal consistency and complementary views. Yet it requires building more advanced algorithms to fully distill information from it. The datasets that we plan to explore are indoor scenes (e.g. ScanNet, RedWood) and outdoor scenes (e.g. Waymo, NuScenes, KITTI), or a subset of them.
In general, our goal is to build machine learning models that generalize well to unseen geometries/scenes. To this end, we will explore modern techniques in self-supervised learning and data augmentation, and tailor them to point cloud sequence data. Students are welcomed to work on any downstream tasks that fit into this framework, e.g., point cloud segmentation/classification, shape generators, object detection and human shape modeling.
3D geometry understanding using point cloud sequences
"More than one model for learning: a computational investigaation"
Faculty Advisor: Professor Robert C. Berwick
Mentor(s):
Contact e-mail: berwick@csail.mit.edu
Research Area(s): Natural Language and Speech Processing
We make decisions in every moment of our lives. Among the very first, and perhaps the most important one, is the language we learn. Every child is capable of learning every language in the world. It is now clear that all languages share many structural regularities.a Universal Grammar.and slo clear that the specific language(s) children learn can only be fixed by input in their linguistic environment.

This project investigates the mechanism of making linguistic decisions: How do children integrate the general constraints on language with the language-specific input? A prominent proposal likens children to "little scientists" who weigh data.the specific linguistic input-against competing scientific hypotheses, selecting the optimal one(s). In a recent paper, the learner is endowed with primitive computational processes akin to string operations. Learning consists of evaluating the combinations of these primitives against the input, a set of examples drawn from a formal language such as a regular languages or context-free languages. The combination(s) with the highest posterior probabilities are selected as winners. A number of important questions, both technical and linguistic, remain; these will be explored in this SuperUrop aiming to find out what computational model best fits what children actually do.
More than one model for learning: a computational investigaation
"Analyzing deterioration of ML models due to distribution shift"
Faculty Advisor: David Sontag
Mentor(s): Christina Ji
Contact e-mail: cji@mit.edu
Research Area(s): Machine Learning
(See more about the Clinical Machine Learning group here: Link )

Machine learning models are sensitive to differences in training and deployment environments. In healthcare, deploying a model in another hospital with different clinical practices, disease trends, or changes in data models may lead to deterioration in clinical decision support tools. Identifying when models have deteriorated and adapting them is important for patient safety. Many transfer learning methods for adapting to new domains have been developed, including importance reweighting, conditional distribution matching, two-stage offset estimation, multi-task learning, and few-shot learning. Other methods, such as invariant risk minimization and group distributionally robust optimization, learn a single model that generalizes across multiple domains.

WILDS is a recently released set of real-world benchmark datasets for distribution shift ( Link ). The paper discusses how the ideal comparison between out-of-distribution loss (from training and testing on different distributions) and in-distribution loss (from both training and testing on the test distribution) is not feasible when there is insufficient test data. Instead, most analyses compare the out-of-distribution loss with the loss from evaluating the same model on held-out training data. While this comparison is feasible, it is not equivalent to the ideal comparison because the test distribution may be harder to model than the training distribution.

To examine how we can get insights about the ideal comparison from this feasible comparison, we are interested in decomposing the difference in the feasible comparison into two components: 1) Bias: How much error is due to selecting the wrong model in the hypothesis class because the model was learned on the training distribution? 2) Irreducible error: How much error is simply because the test distribution is noisier than the training distribution and thus predictions would be worse even if the right model was selected for both distributions? Bias is the difference measured in the ideal comparison, so optimal transfer learning methods would minimize only bias, as irreducible error cannot be addressed.

The primary goal of this SuperUROP is to create a simulator where we can control the amount of bias and irreducible error introduced by the distribution shift. This will allow us to examine when standard transfer learning methods can minimize the introduced bias in two-domain and multi-domain settings. Depending on the UROP student's interests, the project can be extended to developing a method to decompose bias and irreducible error in real-world experiments to determine whether a model is the best possible for a test distribution with limited data. Another potential extension is creating a new transfer learning objective minimizing the bias.

Requirements:
- Earned an A in 6.867, 6.438, 6.437, 6.871, or equivalent grad-level ML or math classes. We may also consider an applicant with an A in 6.036 and significant ML experience from UROPs or internships.
- Proficient in Python
- Pass an ML test during an interview
- Work on the project for at least 12 hours a week
- Commit to the project for the full SuperUROP year
- Motivated and passionate about the project
Analyzing deterioration of ML models due to distribution shift
"More than one model for learning: a computational investigaation"
Faculty Advisor: Professor Robert C. Berwick
Mentor(s):
Contact e-mail: berwick@csail.mit.edu
Research Area(s): Natural Language and Speech Processing
We make decisions in every moment of our lives. Among the very first, and perhaps the most important one, is the language we learn. Every child is capable of learning every language in the world. It is now clear that all languages share many structural regularities.a Universal Grammar.and slo clear that the specific language(s) children learn can only be fixed by input in their linguistic environment.

This project investigates the mechanism of making linguistic decisions: How do children integrate the general constraints on language with the language-specific input? A prominent proposal likens children to "little scientists" who weigh data.the specific linguistic input-against competing scientific hypotheses, selecting the optimal one(s). In a recent paper, the learner is endowed with primitive computational processes akin to string operations. Learning consists of evaluating the combinations of these primitives against the input, a set of examples drawn from a formal language such as a regular languages or context-free languages. The combination(s) with the highest posterior probabilities are selected as winners. A number of important questions, both technical and linguistic, remain; these will be explored in this SuperUrop aiming to find out what computational model best fits what children actually do.
More than one model for learning: a computational investigaation
"GiWATCH: A gastrointestinal mucosa interface for prolonged theragnostics"
Faculty Advisor: Traverso
Mentor(s): Binbin Ying
Contact e-mail: bying@mit.edu
Research Area(s):
The surface of gastrointestinal (GI) tract is covered by mucosal membrane, consisting of enormous health-related biochemical, physiologic, and pathophysiologic information, and serving for nutrition exchange. Progress has been made to access the GI mucosa for diagnostics and therapeutics in clinical settings. However, it is still extremely challenging to build a biocompatible and robust GI mucosa interface enabling real-time, continuous, and minimally invasive interactions with human body, due to the constant GI motility, fast cellular turnover rate, limited cavity space and extremely chemical and biological environments. Here, we aim to explore novel engineering approaches to develop a mucosal platform for long-term deployment of micro-electronics/drug reservoirs/physical barriers in the GI tract. The ultimate applications include but not limit to chronic wound healing, chronic drug delivery, continuous bio-signs monitoring, and closed-loop therapies.
"Programmable smart devices for gastrointestinal-based drug delivery and health monitoring"
Faculty Advisor: Traverso
Mentor(s): Troy Ziliang Kang
Contact e-mail: zkang@mit.edu
Research Area(s):
The gastrointestinal (GI) tract contains all the major organs of the digestive system, making it an essential passageway for chronic, effective and non-invasive drug administration and health monitoring. However, unique characteristics of the GI tract, such as the low PH environment, peristalsis movement of organs, and the mucosa barrier, bring difficulties for the administration of traditional drug-delivery platforms to achieve long retention and localized therapies. While developments of GI-based resident devices are heavily driven by chemical or biological-enabled methods, we are looking for new approaches to program the mechanical responses of the device. Specifically, we are interested in utilizing the environmental stimuli of the GI tract, such as temperature, PH, wetness and etc., to achieve mechanism intelligence---self-actuation and perception of the GI-based device. By participating in this project, you will have opportunities to develop skills in bio-mechanical design, theoretical mechanics, material characterization of smart materials, and ex-vivo and in-vivo experiments. Undergraduates who have backgrounds in mechanical engineering, electrical engineering, material science, bio-engineering and related disciplines are welcome to apply. Programmable smart devices for gastrointestinal-based drug delivery and health monitoring
"Trusted Virtual Machines"
Faculty Advisor: Mengjia Yan
Mentor(s):
Contact e-mail: mengjiay@mit.edu
Research Area(s): Computer Architecture, Computer Systems, Systems (incl OS, databases, computer security)
The project aims to design an open-source Trusted Virtualization Environment on RISC-V processors. The design targets privileged attackers such as the hypervisor of the host operating systems.
The project will involve hardware and software co-design mechanisms to support security features of data encryption, memory management, IO interfaces, etc.

References:
Link
Link
Link

Prerequisites:
- Performed well in 6.004 and 6.033
- Taking 6.858 is a plus
Trusted Virtual Machines
"Exploring Security Vulnerabilities in Modern Processors"
Faculty Advisor: Mengjia Yan
Mentor(s): none
Contact e-mail: mengjia@csail.mit.edu
Research Area(s): Computer Architecture, Computer Systems, Systems (incl OS, databases, computer security)
Modern processors have been aggressively optimized for performance and energy efficiency. However, recent attacks, such as high-profile Spectre and Meltdown attacks, have shown how vulnerable modern computer hardware is.

This projects is to explore various side channel vulnerabilities in modern processors. We are going to study potential security problems in those micro-architecture structures, which have not been fully understood by the community, such as buffers and directories in the memory hierarchy and network-on-chip.

The project will involve reverse engineering commercial processors, developing end-to-end attacks and breaking privacy of modern applications, including crypto libraries and ML applications.

References: Link Link

Prerequisites:
- Performed well in 6.004
- Taking 6.823 is a plus
"Creating apps by speaking in natural language"
Faculty Advisor: Hal Abelson
Mentor(s):
Contact e-mail: hal@mit.edu
Research Area(s): Human Computer Interaction, Machine Learning, Natural Language and Speech Processing
The MIT App Inventor group <appinventor.mit.edu> is dedicated to empowering anyone, even young kids, to create mobile apps for Android and iOS. We're currently creating a system called Aptly, that lets people create apps by describing them at a high level in natural language. You can see a very preliminary demonstration of the system in this short video <Link

The system is based on OpenAI's GPT-3 and Codex and partakes of today's explosion of research in large language models and "no code" development platforms. To work on this project, you should have familiarity with App Inventor, facility in Java and Javascript, and know about current work with GPT-3 and similar large-language models.
Creating apps by speaking in natural language
"Microfluidic experiments for generation of tumor fusion vaccine"
Faculty Advisor: Joel Voldman
Mentor(s): Mahdi Aeinehvand
Contact e-mail: maein@mit.edu,voldman@mit.edu
Research Area(s): BioEECS
A- Overview of cell fusion and microfluidic cell pairing and fusion project:
An exciting area of cancer therapy is developing vaccines comprised of cells that can attack a patient's tumour. These vaccines can be made by fusing a patient's tumor cells with an immune cell. We are developing a method to properly fuse large numbers of cells with unprecedented efficiency, enabling us to target these vaccines to new types of tumors. This is part of a collaboration with Boston-area clinicians.

B- What will a SuperUROP learn and do in this program?
In this program, a SuperUROP will first learn about cell pairing and fusion processes and applications (please see example in publication reference [1]), and then will fabricate microfluidic devices and run cell experiments. The SuperUROP experiments are a mixture of engineering and wet-lab work.

C- We expect that a SuperUROP student who will participate in this (10hr/wk to 12hr/wk) project is willing to:
Take relevant safety and biosafety courses.
Attend training sessions at the beginning of the program to learn about experimental procedure and project objectives.
Fabricate and assemble microfluidic chips.
Participate in various stages of experimentation such as sample prep, microfluidic cell pairing and fusion, and image microscopy.
Attend weekly group and individual meetings, and present activities and progress in bi-weekly schedules.

D- Ref [1] is: Dura, Burak, Yaoping Liu, and Joel Voldman. "Deformability-based microfluidic cell pairing and fusion." Lab on a Chip 14.15 (2014): 2783-2790.

E- Keywords are:
Microfluidic cell pairing and fusion
Soft lithography
Microscopy and imaging
Cancer

Please feel free to contact Mahdi Aeinehvand (maein@mit.edu) and Joel Voldman (voldman@mit.edu) for any further information.
Microfluidic experiments for generation of tumor fusion vaccine
"Bioresorbable Osmotic Pump for Long-term Contraception"
Faculty Advisor: Traverso
Mentor(s): Sanghyun Park
Contact e-mail: sangpark@mit.edu
Research Area(s):
While contraceptive implants are regarded as the most effective reversible option for birth control, the commercial implants are mostly limited to non-bioresorbable matrix-diffusion systems that raise problems like inconstant release rates that compromise their long-term efficacy, the requirement of surgical removal, and a long pharmacokinetic tail that delays the user's return to fertility. Here, we aim to develop a fully bioresorbable osmotic pump system for long-term contraception. Osmotic pump systems are recognized for their capacity to provide near zero-order drug release kinetics coupled with immediate drug release cessation upon completion of osmotic swelling/displacement. In the wet tissue environment, the high osmolality of the osmotic engine causes water flux into the pump through a semi-permeable membrane, and the increased internal pressure from the expansion causes the drug formulation to be pushed out through the orifice on the other side. We ultimately aim to prove the feasibility of the assembled device through pharmacokinetic tests in vivo with rodent and swine models. This fully bioresorbable, zero-order release osmotic pump system has the potential to be an impactful long-term drug delivery platform device for contraception, as well as other diseases that affect the global population where drug adherence is necessary for efficacy. Bioresorbable Osmotic Pump for Long-term Contraception
"Development of off-grid portable desalination unit"
Faculty Advisor: Jongyoon Han
Mentor(s): Dr. Junghyo Yoon
Contact e-mail: jhyoon7@mit.edu
Research Area(s): Applied Physics, Circuits, Energy, Power, Electromagnetics, Environment
Building on the recent progress in the Ion Concentration Polarization (ICP) desalination, we have developed a portable seawater desalination device. The proposed device will be a compact, lightweight electrochemical apparatus capable of removing dissolved and suspended solids in both brackish water and seawater to yield potable water. The device will be useable in remote field settings with good power efficiency and power options. The UROP will develop a controller that enables an operation of two pump and ICP module with user friendly way. We are also working on commercialization.

UROP duty
- Develop controller hardware and software using Python or other languages.
- Test and optimize the controller architecture
- Help prepare for the real-world portable desalination unit deployment.
- (Optional) Engage prototyping of portable desalination unit.

UROP goal
- Learn the state-of-the-art of desalination technology.
- Learn and practice a development of software and hardware for ICP desalination unit.

Additional opportunity
- Opportunity to communicate with industry.
- Opportunities to engage in entrepreneurial activities if interested.

2021 World Water Day Judges Choice: Creative Communication - Portable Desal. Unit For Hydration
Link
Development of off-grid portable desalination unit
"Security Solutions for Anti-Counterfeiting of Agricultural Seeds"
Faculty Advisor: Anantha Chandrakasan, Benedetto Marelli
Mentor(s): Saurav Maji
Contact e-mail: smaji@mit.edu
Research Area(s): Communications, Computer Systems, Signals and Systems
Improved quality of seeds, with techniques like genetic engineering, have helped farmers to improve food productivity. However, there have been an increase in illegal seed practices, including counterfeit seeds, fake seeds, fraudulent labelling, etc. Our group is developing low cost anticounterfeiting solutions by adopting inter-disciplinary approaches from hardware security and materials engineering.

Examples of projects include:

1. Development of a Graphic User Interface Platform for anti-counterfeiting solutions of seeds based on signal-processing algorithms and cryptographic algorithms.
2. Development of machine-learning models for detecting counterfeit seeds.

Prerequisites: Experience with programming in Python & C.
Experience with Image Processing / Machine Learning techniques
Security Solutions for Anti-Counterfeiting of Agricultural Seeds
"Machine Learning, The Cocktail Party Problem, and RF Systems"
Faculty Advisor: Yury Polyanskiy & Greg Wornell
Mentor(s): Gary Lee
Contact e-mail: yp@mit.edu
Research Area(s): Communications, Machine Learning, Signals and Systems
There are many applications where we need to distinguish and separate signals from one another. A classic example arises in audio, where in a room multiple people are speaking at the same time, but there is a need to isolate the audio from a single speaker in such an environment. This is referred to as the "cocktail party problem." Humans can carry out this task surprisingly well, but machines are only beginning to be able to perform as well as humans, aided by powerful new AI and machine learning techniques.

The same problem arises with many other types of signals, including images, video, and, especially, radio frequency signals. For example, both modern wireless devices and sensor systems (such as those being designed for autonomous vehicles) are constantly extracting RF transmissions of interest from among many other transmissions and interfering signals. Here, too, AI and machine learning has the potential to revolutionize how such devices operate, their capabilities, and how we share the RF spectrum more generally.

In this project, you will be involved in the development of AI-based methods for source separation involving RF signals. This will include the use of such techniques in an architecture whereby we first learn signal structure from examples, then carry out the source separation. The techniques to be investigated include, but are not limited to, the use of deep generative models (e.g. GANs, VAEs, Normalizing Flows) and neural network models (RNNs using GRU or LSTM units, U-Nets). Special emphasis is on few-shot and transfer learning approaches (as real-world data is scarce, while synthetic simulated one is abound). The process of prototyping and evaluating the methods arising out of this research will involve using both synthetic data and rich real-world RF recordings.
Machine Learning, The Cocktail Party Problem, and RF Systems
"Autonomous Driving"
Faculty Advisor: Daniela Rus
Mentor(s): none
Contact e-mail: rus@csail.mit.edu
Research Area(s): Cognitive AI, Control and Decision Systems
In this project we are developing perception, planning, control, and machine learning algorithms for autonomous driving. We are emphasizing the capabilities of autonomous vehicles in difficult driving situations such as congestion or weather, and the interactions between autonomous vehicles and human-driven vehicles. We are aiming to develop new algorithms that increase the range of capabilities of the robots and that have provable guarantees.
"Soft Robots"
Faculty Advisor: Daniela Rus
Mentor(s): none
Contact e-mail: rus@csail.mit.edu
Research Area(s): Cognitive AI, Materials, Devices and Photonics, Numerical Methods, Theoretical Computer Science
In this project we address the design, fabrication, and control of a soft robots such as robot fish and robot manipulators. Soft robot generally consist of several segments actuated using bidirectional fluidic elastomer actuators. The robots are fabricated using molding and 3d printing processes. We also develop the associated computation and control systems enable the robots to move.
"Accelerating Compression with DSL"
Faculty Advisor: Vivienne Sze & Saman Amarasinghe
Mentor(s): none
Contact e-mail: sze@mit.edu, samana@mit.edu
Research Area(s): Programming Languages (incl software eng), Signals and Systems
Video is perhaps the biggest of the 'big data' being collected and transmitted. Today, over a billion hours of video are captured every day, and over 70% of Internet traffic is used to transport video. The exponential growth of video places a significant burden on global communication networks. Video compression is critical to keep up with this demand.

Are you interested in accelerating video compression using a domain specific language? This SuperUROP opportunity involves investigating methods to accelerate compression algorithms using a domain specific language we have developed called CoLA. Prerequisites include signal and image processing (6.003, 6.344). Experience with compression algorithms such as JPEG, H.264/AVC, H.265/HEVC, as well as C++ is a plus.
"Efficient Computing for Robotics"
Faculty Advisor: Vivienne Sze & Sertac Karaman
Mentor(s): none
Contact e-mail: sze@mit.edu
Research Area(s): Circuits, Control and Decision Systems, Robotics, Signals and Systems
Autonomous navigation of miniaturized robots (e.g., nano/pico aerial vehicles) is currently a grand challenge for robotics research, due to the need for processing a large amount of sensor data (e.g., camera frames) with limited on-board computational resources. Enabling efficient computing is critical for this task. Our group has been investigating various approaches from both the algorithm and hardware perspective, including both deep neural network (DNN) and non-DNN based solutions.

Low-Energy Autonomy and Navigation Group website: Link

Example projects include
Navion: Link
FastDepth: Link
Robot Exploration: Link

The goal of this project is to contribute to the development of energy-efficient algorithms (visual-inertial navigation, motion planning, mutual-information-based exploration, depth estimation, robot perception, and others) and energy-efficient hardware accelerators for various forms of processing to support real-time navigation. There are several opportunities to get involved in this project ranging from:
** Development of efficient DNN and non-DNN algorithms
** RTL design for the FPGA
** Development embedded software for an ARM core

Prior experience in FPGA design and/or algorithm design is a plus.
Efficient Computing for Robotics

Total: 117

Technical questions related to this site? Please contact Search-Admin@superurop-apply.mit.edu
Accessibility