Events

July 26, 2021

Principled Methods and Models for Deep Learning Based Functional Genomics

Konstantin Krismer

MIT Biological Engineering

10:30A

- 11:30A

Location

A Zoom link will be provided the day before the defense.

Add to Calendar 2021-07-26 10:30:00 2021-07-26 11:30:00 America/New_York Principled Methods and Models for Deep Learning Based Functional Genomics Many advances in functional genomics and in biology more broadly can be attributed to the rise of massively parallel sequencing technology and its derivatives. As the volume of sequencing and other high-throughput experimental data increases exponentially, so does the need for computational methods to analyze and condense these vast amounts of data, and to help explain the underlying phenomena. In this thesis, I describe five projects that introduce novel techniques and methods in functional genomics.The first project introduces a simulation-based framework to investigate neural network architectures that are trained on biological sequence data, as is common in functional genomics. The second project describes a two-pronged approach to study the determinants of cell type-specific chromatin accessibility, with an ensemble of neural networks trained on DNase-seq data to predict chromatin accessibility, and MIAA, the multiplexed integrated accessibility assay, to validate, experimentally, these in silico predictions. The third project presents a method to identify long-range genomic interactions from ChIA-PET and HiChIP data. Enabled by this work, the fourth project aims to provide a means to identify reproducible long-range genomic interactions. We continue the analysis of long-range interactions in the fifth project by performing co-enrichment analysis of transcription factor sequence motifs.Collectively, these methods provide new approaches to a range of problems in functional genomics, from finding appropriate neural network architectures for sequence-based prediction tasks to uncovering patterns in long-range genomic interactions.Thesis Supervisor David K. Gifford, Professor of Electrical Engineering and Computer Science, and Biological Engineering, MITThesis Chair Douglas A. Lauffenburger, Ford Professor of Biological Engineering, Chemical Engineering, and Biology, MIT Thesis Committee Phillip A. Sharp, Institute Professor and Professor of Biology, MIT A Zoom link will be provided the day before the defense.

Algorithms and Hardness for Approximating the Diameter of a Graph

Nicole Wein

MIT CSAIL

3:00P

- 4:00P

Location

https://mit.zoom.us/j/94451184162?pwd=V2tzMXZNUk5xSXJHUHZ3WU5HOUw1UT09 Password: 922944

Zoom

Add to Calendar 2021-07-26 15:00:00 2021-07-26 16:00:00 America/New_York Algorithms and Hardness for Approximating the Diameter of a Graph Abstract: The diameter of a graph is one of the most basic and fundamental attributes of a graph. It is defined as the distance between the pair of vertices that is the farthest apart. The diameter of a graph is a meaningful parameter for many applications such as distributed computation and social networks. We seek fast algorithms for computing the diameter of a graph. This is one of the central problems in the area of fine-grainedcomplexity.The naive algorithm for computing the diameter of a graph is to find the distance between all pairs of vertices and return the largest one. Interestingly, no better algorithm is known. Furthermore, there is evidence from fine-grained complexity, that no subquadratic time algorithm exists for computing the diameter of a graph exactly. In particular, such an algorithm would falsify the Strong Exponential Time Hypothesis (SETH). For applications with very large graphs, even quadratic time can be prohibitively slow. Thus, we turn to approximation algorithms with faster running times. Prior work establishes a hierarchy of algorithms that trade-off time and accuracy, as well as a single lower bound conditioned on SETH.Our first main contribution is the development of a hierarchy of conditional lower bounds under SETH for approximating the diameter of a graph, that establishes a time vs. accuracy trade-off. These lower bounds show that several of the known algorithms on the trade-off curve are conditionally tight.Second, we study the approximability of the diameter of a graph in a variety of natural settings, such as when the graph is changing over time, or when we only care about the distances between particular subsets of vertices, or when we only care about one-way distances in a directed graph. For these variants, we develop both approximation algorithms and conditional lower bounds that are often tight.Thesis Supervisor: Virginia Vassilevska Williams https://mit.zoom.us/j/94451184162?pwd=V2tzMXZNUk5xSXJHUHZ3WU5HOUw1UT09 Password: 922944

July 28, 2021

Normalizing flows: overview and recent advances

George Papamakarios

Google DeepMind

10:00A

- 11:00A

Location

https://mit.zoom.us/j/99352571082

Add to Calendar 2021-07-28 10:00:00 2021-07-28 11:00:00 America/New_York Normalizing flows: overview and recent advances Abstract:Normalizing flows are a set of deep-learning tools for creating flexible and tractable probabilistic models. They can be used for a variety of tasks in probabilistic machine learning, including density estimation, generative modelling, image and audio synthesis, and probabilistic inference. In this talk, I will give an overview of the basics of normalizing flows, present the most common architectures and their computational trade-offs, and discuss some recent advances and latest applications.Bio:George Papamakarios is a Senior Research Scientist at DeepMind, London. He obtained his MSc from Imperial College London and his PhD from the University of Edinburgh. During his PhD he developed a set of new deep-learning methods for density estimation and Bayesian inference. More broadly, his research interests include probabilistic machine learning, normalizing flows, generative modelling, probabilistic inference, and applications to scientific problems. https://mit.zoom.us/j/99352571082

July 30, 2021

Multimodal Representation Learning for Medical Image Analysis

Ruizhi (Ray) Liao

MIT EECS / CSAIL

11:00A

- 12:00P

Location

https://mit.zoom.us/j/95631162606

Add to Calendar 2021-07-30 11:00:00 2021-07-30 12:00:00 America/New_York Multimodal Representation Learning for Medical Image Analysis Thesis supervisor: Polina GollandThesis committee: Peter Szolovits, David SontagMy thesis develops machine learning methods that exploit multimodal clinical data to improve medical image analysis. Medical images capture rich information of a patient’s physiological and disease status, central in clinical practice and research. Computational models, such as artificial neural networks, enable automatic and quantitative medical image analysis, which may offer timely diagnosis in low-resource settings, advance precision medicine, and facilitate large-scale clinical research. Developing such image models demands large training data. Although digital medical images have become increasingly available, limited structured image labels for the image model training have remained a bottleneck. To overcome this challenge, I have built machine learning algorithms for medical image model development by exploiting other clinical data.Clinical data is often multimodal, including images, text (e.g., radiology reports, clinical notes), and numerical signals (e.g., vital signs, laboratory measurements). These multimodal sources of information reflect different yet correlated manifestations of a subject’s underlying physiological processes. I propose machine learning methods that take advantage of the correlations between medical images and other clinical data to yield accurate computer vision models. I use mutual information to capture the correlations and develop novel algorithms for multimodal representation learning by leveraging local data features. The experiments described in this thesis demonstrate the advances of the multimodal learning approaches in the application of chest x-ray analysis. https://mit.zoom.us/j/95631162606

August 03, 2021

Securing Hardware for Designing Trustworthy Systems

Prabhat Mishra

University of Florida

12:00P

- 1:00P

Location

Add to Calendar 2021-08-03 12:00:00 2021-08-03 13:00:00 America/New_York Securing Hardware for Designing Trustworthy Systems Abstract: System-on-Chip (SoC) is the brain behind computing and communication in a wide variety of systems, starting from simple electronic devices in smart homes to complex navigation systems in airplanes. Reusable hardware Intellectual Property (IP) based SoC design has emerged as a pervasive design practice in the industry to dramatically reduce SoC design and verification cost while meeting aggressive time-to-market constraints. Growing reliance on these pre-verified hardware IPs, often gathered from untrusted third-party vendors, severely affects the security and trustworthiness of computing platforms. These IPs may come with deliberate malicious implants to incorporate undesired functionality, undocumented test/debug interface working as hidden backdoor, or other integrity issues. It is crucial to evaluate the integrity and trustworthiness of third-party IPs for designing trustworthy systems. In this talk, I will introduce a wide variety of hardware security vulnerabilities, design-for-security solutions, and possible attacks and countermeasures. I will briefly describe how the complementary abilities of simulation-based validation, formal verification as well as side channel analysis can be effectively utilized for comprehensive SoC security and trust validation. I will conclude with a discussion on application-specific security solutions as well as future hardware security challenges.Biography: Prabhat Mishra is a Professor in the Department of Computer and Information Science and Engineering and a UF Research Foundation Professor at the University of Florida. He received his Ph.D. in Computer Science from the University of California at Irvine in 2004. His research interests include embedded and cyber-physical systems, hardware security and trust, computer architecture, energy-aware computing, formal verification, system-on-chip validation, machine learning, and quantum computing. He has published 8 books, 35 book chapters, 16 patents/copyrights, and more than 200 research articles in premier international journals and conferences. His research has been recognized by several awards including the NSF CAREER Award, IBM Faculty Award, ten best paper awards and nominations, and EDAA Outstanding Dissertation Award. He currently serves as an Associate Editor of IEEE Transactions on VLSI Systems. He is an IEEE Fellow and an ACM Distinguished Scientist, and served as an ACM Distinguished Speaker during 2016-2019.Zoom:Join Zoom Meetinghttps://mit.zoom.us/j/97527284254Password: <3securityOne tap mobile+16465588656,,97527284254# US (New York)+16699006833,,97527284254# US (San Jose)Meeting ID: 975 2728 4254US : +1 646 558 8656 or +1 669 900 6833International Numbers: https://mit.zoom.us/u/auBvg4NEVJoin by SIP97527284254@zoomcrc.comJoin by Skype for Businesshttps://mit.zoom.us/skype/97527284254

August 23, 2021

Efficient Deep Learning: From Theory to Practice

Lucas Liebenwein

MIT CSAIL

3:00P

- 4:30P

Location

https://mit.zoom.us/j/98186493200?pwd=Z0hTZVMyWWFVUTl3ckd2cHpHRFg5Zz09 or 32-G449 (Patil/Kiva)

Add to Calendar 2021-08-23 15:00:00 2021-08-23 16:30:00 America/New_York Efficient Deep Learning: From Theory to Practice Abstract:Modern machine learning often relies on deep neural networks that are prohibitively expensive in terms of the memory and computational footprint. This in turn significantly inhibits the potential range of applications where we are faced with non-negligible resource constraints, e.g., real-time data processing, embedded devices, and robotics. In this thesis, we develop theoretically-grounded algorithms to reduce the size and inference cost of modern large-scale neural networks. By taking a theoretical approach from first principles, we intend to understand and analytically describe the performance-size trade-offs of modern deep networks, i.e, the generalization properties, and to leverage such insights to devise practical algorithms for obtaining efficient neural network models via pruning or compression. Beyond theoretical aspects and the inference time efficiency of neural networks, we study how compression can yield novel insights into the design and training of neural networks. We investigate the practical aspects of the generalization properties of pruned neural network beyond simple metrics such as test accuracy. We then show how in certain applications pruned neural networks can improve the training and hence generalization of networks.Thesis supervisor: Daniela RusThesis committee: Michael Carbin, Song Han https://mit.zoom.us/j/98186493200?pwd=Z0hTZVMyWWFVUTl3ckd2cHpHRFg5Zz09 or 32-G449 (Patil/Kiva)

September 09, 2021

Implementing Symbols and Rules with Neural Networks

Ellie Pavlick

Brown University and Google

7:00P

- 8:00P

Location

Virtual: please register here: https://acm-org.zoom.us/webinar/register/6016263861654/WN_kBPd0z0MR7CDqJy26Gj0Mw

Add to Calendar 2021-09-09 19:00:00 2021-09-09 20:00:00 America/New_York Implementing Symbols and Rules with Neural Networks IEEE Computer Society and GBC/ACMonline 7:00 PM, Thursday, 9 September 2021Implementing Symbols and Rules with Neural NetworksEllie PavlickRegister in advance for this webinar athttps://acm-org.zoom.us/webinar/register/6016263861654/WN_kBPd0z0MR7CDqJy26Gj0MwAfter registering, you will receive a confirmation email containing information about joining the webinar.Many aspects of human language and reasoning are well explained in terms of symbols and rules. However, state-of-the-art computational models are based on large neural networks which lack explicit symbolic representations of the type frequently used in cognitive theories. One response has been the development of neuro-symbolic models which introduce explicit representations of symbols into neural network architectures or loss functions. In terms of Marr's levels of analysis, such approaches achieve symbolic reasoning at the computational level ("what the system does and why") by introducing symbols and rules at the implementation and algorithmic levels. In this talk, I will consider an alternative: can neural networks (without any explicit symbolic components) nonetheless implement symbolic reasoning at the computational level? I will describe several diagnostic tests of "symbolic" and "rule-governed" behavior and use these tests to analyze neural models of visual and language processing. Our results show that on many counts, neural models appear to encode symbol-like concepts (e.g., conceptual representations that are abstract, systematic, and modular), but not perfectly so. Analysis of the failure cases reveals that future work is needed on methodological tools for analyzing neural networks, as well as refinement of models of hybrid neuro-symbolic reasoning in humans, in order to determine whether neural networks' deviations from the symbolic paradigm are a feature or a bug.Ellie Pavlick is the Manning Assistant Professor of Computer Science at Brown University and a Research Scientist at Google AI.Ellie received her PhD in Computer and Information Science from University of Pennsylvania in 2017. In 2012, she received a Bachelor of Arts in Economics from Johns Hopkins University and a Bachelor of Music in Saxophone Performance from the Peabody Conservatory. Ellie's current research is in Natural Language Processing, specifically on computational models of semantics and pragmatics which emulate human inferences. She is interested in building better computational models of natural language semantics and pragmatics: how does language work, and how can we get computers to understand it the way humans do?This joint meeting of the Boston Chapter of the IEEE Computer Society and GBC/ACM will be online only due to the COVID-19 lockdown.Up-to-date information about this and other talks is available online at https://ewh.ieee.org/r1/boston/computer/. You can sign up to receive updated status information about this talk and informational emails about future talks at https://mailman.mit.edu/mailman/listinfo/ieee-cs, our self-administered mailing list. Virtual: please register here: https://acm-org.zoom.us/webinar/register/6016263861654/WN_kBPd0z0MR7CDqJy26Gj0Mw