Add to Calendar
2025-12-17 13:00:00
2025-12-17 14:00:00
America/New_York
PhD defense, Rickard Brüel Gabrielsson: Feature Learning for Foundation Models Across Tasks, Modalities, and Scales
TBD
December 17
December 11
Towards Interpretable and Operationalized Fairness in Machine Learning
Schrasing Tong
MIT CSAIL
Add to Calendar
2025-12-11 15:30:00
2025-12-11 16:30:00
America/New_York
Towards Interpretable and Operationalized Fairness in Machine Learning
Thesis advisor: Lalana KagalThesis committee: Peter Szolovits, Brian HeddenAbstractMachine learning systems are increasingly deployed in sensitive, real-world settings, yet persistent biases in model predictions continue to disadvantage marginalized groups. This thesis develops practical and interpretable methods for understanding and mitigating such biases in natural language generation and computer vision. For large language models, we introduce a decoding-time approach that leverages small biased and anti-biased expert models to obtain a debiasing signal that is added to the LLM output. This approach combines computational efficiency - fine-tuning a small model versus re-training a large model and interpretability - one can examine the probability shift from debiasing. In computer vision, we leverage concept bottleneck models (CBMs), which map images to human-understandable concepts, to improve transparency and help mask proxy features that correlate with sensitive attributes. To counter CBM information leakage and improve fairness-performance tradeoffs, we introduce three mitigation strategies: (1) reducing leakage with a top-k concept filter, (2) removing concepts that correlate strongly with gender, and (3) applying adversarial debiasing to further suppress sensitive information. Together, these contributions illustrate how interpretability and operationalization can make fairness interventions more trustworthy, scalable, and aligned with real deployment needs.
TBD
December 09
[Thesis Defense] Yung-Sung Chuang: "Towards Factual and Trustworthy Large Language Models"
MIT CSAIL
Add to Calendar
2025-12-09 15:00:00
2025-12-09 16:00:00
America/New_York
[Thesis Defense] Yung-Sung Chuang: "Towards Factual and Trustworthy Large Language Models"
Thesis Advisor: James GlassThesis Committee: Yoon Kim, Jacob AndreasCalendar Invitation: http://people.csail.mit.edu/yungsung/defense.icsSpeaker's Website: https://yung-sung.github.ioAbstract: Large Language Models (LLMs) have transformed how we interact with information, yet hallucinations, e.g., plausible but factually incorrect outputs, remain a critical barrier to their deployment in high-stakes applications. This thesis presents a comprehensive approach to understanding and mitigating hallucinations across several fundamental dimensions of knowledge in AI systems: parametric, contextual, and attribution knowledge.We identify that hallucinations arise from different failure modes requiring distinct solutions. First, models may fail to leverage parametric knowledge already encoded in their weights. We introduce DoLa (Decoding by Contrasting Layers), which amplifies factual knowledge by dynamically contrasting predictions across transformer layers, improving factuality without training or external knowledge. Second, in retrieval-augmented generation settings, models often fail to properly use provided context. We develop Lookback Lens, which analyzes attention patterns to detect and reduce hallucinations. Third, even when models generate correct content, users need verifiable evidence. We present SelfCite, a self-supervised alignment method that enables LLMs to provide accurate sentence-level citations through a reward design of context ablation. Together, these methods form a roadmap towards better AI systems, working towards systems that are not only capable but also reliable, transparent, and trustworthy.
TBD
December 01
Toward Provable Privacy for Black-Box Algorithms via Algorithmic Stability
Mayuri Sridhar
MIT CSAIL
Add to Calendar
2025-12-01 13:00:00
2025-12-01 14:30:00
America/New_York
Toward Provable Privacy for Black-Box Algorithms via Algorithmic Stability
This thesis focuses on enabling the design of algorithms with provable privacy as a first-order goal. We focus on the following setting: an algorithm is trained on sensitive data and the result is then exposed for public use -- how can we quantify the privacy risk of this exposure? Prior work typically focuses on providing privacy through privatizing specific algorithms. These techniques have two main drawbacks: (1) they require significant white-boxing per algorithm and (2) the privacy-utility tradeoffs may be hard to quantify.We first leverage the PAC privacy framework to mitigate the white-boxing requirements. In this talk, we show how we can privatize a wide range of database queries in a black-box manner. We discuss how to build a simple privatization layer, PAC-DB, that can provide provable privacy guarantees for general SQL queries. This allows us to expand the capabilities of private database analytics, enabling complex queries without the use of a trusted curator.We then focus on understanding the utility impacts of privatization. We focus on designing privacy-conscious algorithms. That is, rather than first constructing an algorithm then computing the noise required to privatize it --- a paradigm we refer to as post-hoc privatization --- we optimize the algorithm's hyperparameters given the privacy budget. We instantiate this via regularized linear regression. In particular, we derive the theoretically-optimal regularization weight to maximize utility under a provided privacy budget. We provide experimental results showing the benefits of privacy-conscious design over post-hoc privatization.
TBD
October 24
Add to Calendar
2025-10-24 14:00:00
2025-10-24 15:00:00
America/New_York
(THESIS DEFENSE) LUJIE YANG: "Bridging Model-Based and Learning-Based Methods for Robotic Loco-Manipulation and Control"
Speaker: Lujie YangHost: Prof. Russ TedrakeDate: Friday - October 24, 2025Time: 2:00 - 3:00pm ETLocation: Room 32-G449 (Patil/Kiva)Zoom Link: https://mit.zoom.us/j/95546717935 Abstract: Learning-based neural network (NN) control policies have shown impressive empirical performance in a wide range of tasks in robotics and control, including autonomous driving, drone racing, locomotion and manipulation. Yet, these policies remain constrained by their dependence on large quantities of high-quality data and their inability to provide the performance guarantees demanded in safety-critical settings. In contrast, model-based optimization methods exploit system structures to provide strong guarantees and enable efficient offline computation, but they often require accurate models and struggle with complex, long-horizon tasks. This thesis bridges the gap between these paradigms by leveraging model-based tools to address the limitations of learning-based methods. First, I introduce a scalable framework for physics-driven data generation, combining human demonstrations with trajectory optimization to produce large-scale, dynamically feasible datasets for dexterous manipulation. Behavior cloning policies trained on these datasets achieve robust performance and generalization across different embodiments and physical parameters. I further extend this framework to whole-body loco-manipulation by integrating human-object interaction demonstrations with interaction-preserving optimization to generate kinematically consistent motions across diverse spatial configurations. These synthesized trajectories are then used to guide reinforcement learning to produce dynamically consistent policies with improved robustness and generalization.Finally, to equip learned policies with formal guarantees, I propose an optimization-inspired approach for synthesizing and verifying NN controllers with Lyapunov stability guarantees, scalable to partially observable systems via GPU-accelerated verification. Together, these contributions integrate model-based structure with learning-based flexibility, advancing data-efficient, generalizable, and verifiable robotic control.
TBD
August 08
Add to Calendar
2025-08-08 14:30:00
2025-08-08 16:00:00
America/New_York
UPDATED: Room Change [Thesis Defense - Rahul Ilango] Metacomplexity and Metacryptography
Abstract:In this thesis, we show --- for the first time --- how to achieve two longstanding dream results in theoretical computer science, under common assumptions. First, we determine the exact computational complexity of finding optimal circuits. Specifically, we show that the Minimum Circuit Size Problem is NP-complete, a dream result since Levin defined NP-completeness in 1973. Second, we construct traditional mathematical proofs that reveal nothing other than the truth of the statement being proven. Specifically, we show that zero-knowledge proofs for NP with perfect soundness and truly no interaction are effectively possible, a dream result since Goldwasser, Micali, and Rackoff defined zero-knowledge in 1985. A core part of both results is a "meta'' perspective on the areas of complexity theory and cryptography.
TBD
July 25
Add to Calendar
2025-07-25 13:00:00
2025-07-25 14:00:00
America/New_York
Thesis Defense: Programmable Architectural Support for Diverse Sparse Workloads - Ryan Lee
Defense Title: Programmable Architectural Support for Diverse Sparse WorkloadsSparsity is abundant in many workload domains, but presents challenges that results in under- utilization of the available resources in existing hardware. Sparse workloads exhibit irregular control-flow and long-latency memory accesses, starving the core of useful work, and perform fine-grained accesses leading to inefficient use of the available memory bandwidth.Prior work has proposed several software and hardware mechanisms to accelerate sparse workloads, but there has been a lack of a general technique that is applicable to the diverse set of applications in this domain. In particular, existing solutions have had limited support for workloads that concurrently read and update the underlying sparse data structure, such as dynamic graph applications and databases. Prior proposals have instead limited various dimensions of the applications they target in this space, such as restricting the formats they support (e.g., only hash tables) or constraining the types of concurrent operations (e.g., read- only), thereby limiting their applicability. In addition, prior work has insufficiently addressed the inefficient data transfer between compute and memory, instead opting to put expensive compute elements near memory or only supporting restricted forms of fine-grained accesses.This thesis shows that it is possible to design a general and programmable architecture that supports a wide range of sparse workloads. To this end, this thesis presents two hardware accelerators. First, Terminus adds a small hardware unit near each core that accelerate a wide range of data structures types and concurrent reads and updates to these structures, achieving a gmean of 7.4× speedup over a CPU baseline. Second, Gist enhances each DRAM chip with a flexible hardware unit that autonomously performs fine-grained scatter/gather operations for sparse workloads. This allows Gist to more efficiently use the memory bus by returning a compact stream of data, and achieves a gmean of 1.6× speedup over state-of-the-art support for sparse workloads.https://mit.zoom.us/j/8203717891Advisor: Professor Daniel Sanchez
TBD
July 18
Thesis Defense: Optimizing Data Layouts for Evolving Cloud Table Storage
Siva Sudhir
MIT CSAIL
Add to Calendar
2025-07-18 14:00:00
2025-07-18 15:00:00
America/New_York
Thesis Defense: Optimizing Data Layouts for Evolving Cloud Table Storage
Modern data analytics platforms increasingly adopt disaggregated architectures, storing data in cost-effective cloud object stores. While this approach enables a clean separation of concerns, allowing each layer to be independently managed and scaled, it introduces significant performance bottlenecks due to expensive data movement. Effective data layouts, which organize data to minimize unnecessary data reads, are thus critical to achieving high query performance. However, existing techniques typically rely on manually specified layouts, collect limited metadata, or lack mechanisms to dynamically adapt to changing data and workloads.This thesis investigates adaptive, metadata-rich, expressive data layouts for cloud table storage. First, we introduce Pando, a correlation-aware layout technique that leverages rich metadata on query predicates to significantly improve data skipping. Next, we propose CopyRight, a partial replication strategy that selectively replicates subsets of data and optimizes each replica differently, efficiently serving heterogeneous query patterns. Finally, we describe Self-Organizing Data Containers (SDCs), a practical table storage layer for the cloud that incrementally reorganizes complex data layouts based on changes in data and workload distributions.-- Please email siva@csail.mit.edu or markakis@mit.edu for the Zoom password.
TBD
July 15
Add to Calendar
2025-07-15 14:00:00
2025-07-15 15:00:00
America/New_York
THESIS DEFENSE: Seeing Beyond Limits with Physics-Informed Priors
THESIS DEFENSE: Seeing Beyond Limits with Physics-Informed PriorsSpeaker: Yang LiuSpeaker Affiliation: MIT EECS & CSAILHost: Frédo DurandHost Affiliation: MIT EECS & CSAILDate: Tuesday, July 15, 2025Time: 2:00 PM to 3:00 PMLocation: 32-D463 (Star) or Zoom Link: https://mit.zoom.us/j/98534109114Abstract:Conventional imaging systems face inherent dimensionality and visibility limits, primarily because image sensors are typically two-dimensional, and light tends to diffuse on rough surfaces or scatter within complex media. In this talk, I will reframe imaging systems through the lens of optical encoding and neural decoding, presenting my key contributions aimed at transcending the traditional limits of dimensionality and visibility. The idea is modelling the forward physical process and iteratively optimizing it with deep denoisers as visual priors, where eventually the priors are physics-informed. First, I introduce Privacy Dual Imaging, which reveals the privacy risk that ambient light sensors embedded in most smart devices could capture images of the scene in front of the screen. This idea of seeing the invisible from subtle intensity fluctuations is inspired by George Orwell’s novel 1984, wherein Big Brother is watching you through a two-way telescreen, and it closely relates to incoherent lensless imaging and non-line-of-sight imaging. Second, I present Snapshot Compressive Imaging, which encodes multiple temporal, spectral, or angular frames into a single measurement captured by a standard two-dimensional sensor. By learning high-dimensional visual priors from image or video data, we can efficiently reconstruct the original higher-dimensional data cube at scale. Lastly, I show that large AI models, particularly diffusion models, can serve as generic visual priors for both cases and beyond. I aim to push the boundaries of imaging and sensing within relevant domains of AI for science and healthcare (with an example).Committee Members: Frédo Durand (advisor, MIT), William T. Freeman (MIT & Google), Kaiming He (MIT & Google)Relevant URL: https://mit.zoom.us/j/98534109114For more information please contact: Roger White <whiter@mit.edu>
TBD
June 02
Add to Calendar
2025-06-02 9:00:00
2025-06-02 11:00:00
America/New_York
[Thesis Defense] Personalizing Robot Assistance under Uncertainty about the Human
Date: June 2Time: 9:00-11:00 AM ETLocation: 32-155Zoom: https://mit.zoom.us/j/9731989629Title: Personalizing Robot Assistance under Uncertainty about the HumanAbstractRobots have the potential to improve the quality of life by assisting with daily tasks, such as helping older adults and people with disabilities get dressed. But meaningful assistance requires personalization: each person has unique preferences, behaviors, and needs.A central challenge is that robots often operate under uncertainty about the human they are helping. This uncertainty may involve the person's preferences, hidden physical states, or reactions to assistance. If not properly addressed, such uncertainty can lead to ineffective, undesired, or even unsafe outcomes.This thesis asks: How should a robot behave when it is uncertain about the human? I present a unified framework for uncertainty-aware personalization in human-robot interaction, spanning three core components of robot intelligence: preference learning, state estimation, and motion planning.1. Preference learning: I introduce the first method that uses response time, a subtle but informative cognitive signal, as implicit feedback. By combining human choices with response times, robots can infer not only what a person prefers but also how strongly they prefer it. This reduces uncertainty and accelerates preference learning.2. State estimation: To support safe physical assistance when parts of the human body (e.g., the elbow) are occluded, I introduce a state estimator that models uncertainty in learned human dynamics and robot sensing. It constructs a geometric set (e.g., a 3D box) that reliably contains the true hidden human state, enabling safer and more precise robot behavior.3. Motion planning: When a robot is uncertain about future human motion, it may behave overly conservatively to avoid causing harm, resulting in ineffective assistance. To address this, I propose a relaxed safety formulation that allows the robot to either avoid collisions or make low-impact contact. This approach enables the robot to act more effectively while still maintaining safety under uncertainty.Together, these contributions lay a foundation for assistive robots that personalize their behavior while adapting to the uncertain and dynamic nature of human needs.Thesis Supervisor: Julie A. ShahCommittee Members: Julie A. Shah, Dylan Hadfield-Menell, Na (Lina) Li, Aude BillardThesis Readers: Vaibhav Unhelkar, Tariq IqbalContact: shenli@mit.edu
TBD
May 29
Add to Calendar
2025-05-29 10:00:00
2025-05-29 12:00:00
America/New_York
[Thesis Defense] Generalizable Robot Manipulation through Unified Perception, Policy Learning, and Planning
Abstract:Advancing robotic manipulation to achieve generalization across diverse goals, environments, and embodiments is a critical challenge in robotics research. While the availability of data and large-scale training has brought exciting progress in robotics manipulation, current methods often struggle with generalizing to unseen, unstructured environments and solving long-horizon tasks. In this thesis, I will present my contributions that bridge structured decision-making frameworks with learned perceptual and policy components to enable multi-step manipulation in partially observable environments. Specifically, I will talk about my work in 1) constructing a modular framework that estimates affordances using learned perceptual models with task and motion planning (TAMP) for object rearrangement in unstructured scenes, 2) learning generative diffusion models of robot skills, which can be composed to solve unseen combination of environmental constraints through infeference-time optimization, 3) leveraging large vision-language models (VLMs) in building task-oriented visual abstractions, allowing skills to generalize across different environments with only 5 to 10 demonstrations. Together, these approaches contribute to the generality and scalability of embodied agents towards solving real-world manipulation in unstructured environments.Thesis Committee: Leslie Kaelbling, Tomás Lozano-Pérez, Russ Tedrake
TBD
May 12
Thesis Defense: Scaling Cooperative Intelligence via Inverse Planning and Probabilistic Programming
MIT
Add to Calendar
2025-05-12 11:00:00
2025-05-12 12:30:00
America/New_York
Thesis Defense: Scaling Cooperative Intelligence via Inverse Planning and Probabilistic Programming
Thesis Defense: Scaling Cooperative Intelligence via Inverse Planning and Probabilistic ProgrammingPresenter: Tan Zhi-XuanPlease email xuan@mit.edu for Zoom linkHow can we build cooperative machines that model and understand human minds — machines that assist us with our goals, coordinate on shared plans, infer the intentions behind our words, and even learn our norms and values? In this talk, I will introduce a scalable Bayesian approach to building such systems via inverse planning and probabilistic programming. By combining online model-based planners and sequential Monte Carlo inference into a single architecture, Sequential Inverse Plan Search (SIPS), we can infer human goals from actions in faster-than-real-time, while scaling to environments with hundreds of possible goals and long planning horizons that have proved intractable for earlier methods. SIPS can additionally make use of large language models (LLMs) as likelihood functions within probabilistic programs, allowing us to build AI assistants and copilots that reliably infer human goals from ambiguous instructions, then provide assistance under uncertainty with much higher success rates than LLMs can on their own. By applying this Bayesian approach in many-agent environments, we are also able to design agents that rapidly learn cooperative social norms from others' behavior, achieving mutually beneficial outcomes with orders of magnitude less data than model-free deep RL. I will conclude by charting out how this research program could deliver a new generation of cooperative AI systems grounded in rational AI engineering, while illuminating the computational foundations of human cooperation and addressing fundamental challenges in building human-aligned AI.Thesis Committee: Vikash Mansinghka, Joshua Tenenbaum, Dylan Hadfield-Menell, Leslie Kaelbling
TBD
May 08
Mark Hamilton's Thesis Defense - Unsupervised Discovery of Structure in Complex Systems
MIT and Microsoft
Add to Calendar
2025-05-08 15:00:00
2025-05-08 16:00:00
America/New_York
Mark Hamilton's Thesis Defense - Unsupervised Discovery of Structure in Complex Systems
How does the human mind make sense of raw information without being taught how to see or hear? This thesis will explore how to build algorithms that can uncover interpretable structure from large collections of data like images and video without needing human annotations or labels. First, we will see how to build algorithms that can perform tasks like classifying every pixel of the world, localizing sound, and decoding natural language, just by watching unlabeled videos without any knowledge of text. Second, we will see how these ideas lead us to a new unifying theory of representation learning. In particular, I will show how 20+ common machine learning methods, such as dimensionality reduction, clustering, contrastive learning, and spectral methods emerge from a single unified equation. Finally, we will discuss how this unifying theory applies to our ongoing efforts to decode animal communication using large-scale, unsupervised, and interpretable learners. We will conclude with some preliminary analysis of the complex vocalizations of Atlantic Spotted Dolphins.
TBD
Add to Calendar
2025-05-08 14:15:00
2025-05-08 15:30:00
America/New_York
Automatic Integration and Differentiation of Probabilistic Programs
Automatic Integration and Differentiation of Probabilistic ProgramsPresenter: Alexander LewThesis Supervisors: Vikash K. Mansinghka and Joshua B. TenenbaumDate: May 8, 2025Time: 2:15pm ETLocation: Building 46, room 3310Please contact alexlew@mit.edu for a Zoom link.Abstract:By automating the error-prone math behind deep learning, systems such as TensorFlow and PyTorch have supercharged machine learning research, empowering hundreds of thousands of practitioners to rapidly explore the design space of neural network architectures and training algorithms. In this talk, I will show how new programming language techniques, especially generalizations of automatic differentiation, make it possible to generalize and extend such systems to support probabilistic models. Our automation is implemented as a suite of composable program transformations for integrating, differentiating, and deriving densities of probabilistic programs. These transformations are rigorously proven sound using new semantic techniques for reasoning about expressive probabilistic programs, and static types are employed to ensure important preconditions for soundness, eliminating large classes of implementation bugs. Providing a further boost, our tools can help users correctly implement fast, low-variance, unbiased estimators of integrals, gradients, and probability densities that are too expensive to compute exactly, enabling orders-of-magnitude speedups in downstream optimization and inference algorithms.To illustrate the value of these techniques, I’ll show how they have helped us experiment with new architectures that could address key challenges with today’s dominant AI models. In particular, I’ll showcase systems we’ve built for (1) auditable reasoning and learning in relational domains, enabling the detection of thousands of errors across millions of Medicare records, and (2) probabilistic inference over large language models, enabling small open models to outperform GPT-4 on several code generation and constrained generation benchmarks.
TBD
May 07
Add to Calendar
2025-05-07 14:15:00
2025-05-07 15:15:00
America/New_York
Thesis Defense, Noah Golowich - Title: Theoretical Foundations for Learning in Games and Decision-Making
Abstract: As learning algorithms become increasingly capable of acting autonomously, it is important to better understand the behavior that results from their interactions (1) amongst themselves and (2) with their environments. This talk will present work addressing each of these aspects:(1) A pervasive challenge in multi-agent learning settings, which spans both theory and practice and dates back decades, has been the failure of convergence for iterative algorithms such as gradient descent. Accordingly, a longstanding central question with broad relevance is: how quickly can we compute solution concepts, i.e., equilibria, in multi-agent settings? I will discuss results which address this question at several scales, starting with simpler normal-form games and building up to larger games such as extensive-form games.(2) To understand how agents can optimally act in dynamic environments, the framework of reinforcement learning (RL) is used. A notorious challenge in RL is partial observability of the environment, which is typically modeled using Partially Observable Markov Decision Processes (POMDPs). Many existing provable guarantees for POMDPs relied on computationally intractable oracles. I will present the first guarantees for end-to-end learning of a near-optimal policy under a simple condition on the environment known as observability.
TBD
May 02
Add to Calendar
2025-05-02 15:30:00
2025-05-02 17:30:00
America/New_York
[Thesis Defense - Ce Jin] Exploiting Additive Structure in Algorithm Design and Fine-Grained Complexity
Abstract:In this thesis, we investigate the fine-grained complexity of various algorithmic problems with an additive flavor, including 3SUM, Subset Sum, and their close relatives. We explore their connections to various areas, such as graph algorithms, discrete optimization, combinatorial pattern matching, and computational geometry. Our new results include improved algorithms and conditional lower bounds for a wide range of problems, answering multiple open questions from the literature:-Conditional lower bounds for graph problems: We prove new lower bounds for 4-Cycle Listing and Approximate Distance Oracles conditioned on the 3SUM Hypothesis. As a key intermediate step, we show a fine-grained reduction from 3SUM to the special case of 3SUM where all pairwise sums of input numbers are distinct.-Combinatorial pattern matching: We design improved algorithms for Text-to-Pattern Hamming Distances, Pattern Matching with Wildcards, and Geometric Pattern Matching, by drawing connections from 3SUM and sparse convolution.-Knapsack-type problems: We obtain a pseudo-polynomial time algorithm for 0-1 Knapsack with (conditionally) near-optimal dependence on the maximum item weight, an improved approximation scheme for the counting problem #Knapsack, and improved exponential time algorithms for the total search problem Pigeonhole Equal Subset Sum.In order to obtain these results, we employ and develop techniques based on convolution algorithms and their extensions, as well as classic tools from additive combinatorics. Thesis Committee: Ryan Williams (advisor), Virginia Vassilevska Williams (advisor), and Mohsen Ghaffari
TBD
Add to Calendar
2025-05-02 9:30:00
2025-05-02 10:30:00
America/New_York
(Thesis Defense) Building Intelligence that can Interact with the Physical World
Speaker: Johnson Tsun-Hsuan WangAffiliation: MIT EECS (CSAIL)Title: [Thesis Defense] Building Intelligence that can Interact with the Physical WorldDate: Friday, May 2nd 2025Time: 9:30 am EDTLocation: 32-G449 (Patil/Kiva)Zoom: https://mit.zoom.us/j/95448197150?pwd=W0uEtKXgUjoXawXp2cGWrIcsFGtGlO.1Abstract: Recent advances in Artificial Intelligence (AI) have demonstrated remarkable success in parsing, reasoning, and generating digital content across modalities such as natural language, speech, images, videos, and 3D data. However, these breakthroughs have yet to extend meaningfully beyond the digital realm into the physical world. Developing AI for physical interaction poses challenges such as limited grounding, scarce physical data, and high reliability demands in safety-critical settings.This talk outlines a holistic approach to physical AI—through the lenses of data, brain, and body. We begin with data, the foundation of learning, and introduce data-driven and knowledge-driven robot simulation that generates data to improve policy learning and to systematically evaluate and probe existing models. Next, we turn to the brain, focusing on how to bridge the internet-scale knowledge of digital AI with the physical world to improve generalization and interpretability. Finally, we examine the body—the morphological component of intelligence—demonstrating how pre-trained generative models, when integrated with physics-based simulation, can automate the design of robot bodies. Together, this talk explores how digital AI can be extended into the physical world through a comprehensive investigation of data, brain, and body – laying the groundwork for building physical AI.Committee:Prof. Daniela Rus, MIT CSAIL (Advisor)Prof. Sertac Karaman, MIT LIDSProf. Wojciech Matusik, MIT CSAIL
TBD
February 25
[Thesis Defense] Steering Robots with Inference-Time Interactions
Felix Yanwei Wang
EECS/CSAIL
Add to Calendar
2025-02-25 12:00:00
2025-02-25 13:30:00
America/New_York
[Thesis Defense] Steering Robots with Inference-Time Interactions
Date: Tuesday, February 25, 2025Time: 12:00 PM - 1:30 PMLocation: 45-792Zoom: https://mit.zoom.us/j/95052951960Abstract:Imitation learning has driven the development of generalist policies capable of autonomously solving multiple tasks. However, when a pretrained policy makes errors during deployment, there are limited mechanisms for users to steer its behavior. While collecting additional data for fine-tuning can address such issues, doing so for each downstream use case is inefficient at scale. My research proposes an alternative perspective: framing policy errors as task misspecifications rather than skill deficiencies. By enabling users to specify tasks unambiguously via interactions at inference-time, the appropriate skill for a given context can be retrieved without fine-tuning. Specifically, I propose (1) inference-time steering, which leverages human interactions for single-step task specification, and (2) task and motion imitation, which uses symbolic plans for multi-step task specification. These frameworks correct misaligned policy predictions without requiring additional training, maximizing the utility of pretrained models while achieving inference-time user objectives.Thesis Supervisor: Julie ShahCommittee Members: Leslie Kaelbling, Jacob Andreas, Dorsa SadighContact: felixw@mit.edu
TBD
February 05
Add to Calendar
2025-02-05 15:00:00
2025-02-05 16:30:00
America/New_York
Thesis Defense: Designing Hardware Accelerators for Solving Sparse Linear Systems - Axel Feldmann
Solving systems of linear equations with sparse coefficient matrices is a key primitive that sits at the heart of many important numeric algorithms. Because of this primitive's importance, algorithm designers have spent many decades optimizing linear solvers for high performance hardware. However, despite their efforts, existing hardware has let them down. State-of-the-art linear solvers often utilize <1% of available compute throughput on existing architectures such as CPUs and GPUs.There are many different algorithms used to solve sparse linear systems. These algorithms are diverse and often have very different computational bottlenecks. These include low arithmetic intensity, fine-grained parallellism, common control dependences, and sparsity-induced load imbalance.This thesis studies the problem of designing hardware accelerators for sparse linear solvers. We propose three novel architectures that explore different parts of the design space. First, we introduce Spatula, an architecture designed to accelerate direct solvers. Then, we propose Azul, a hardware accelerator targeted at iterative solvers. Taken together, Spatula and Azul demonstrate significant speedups on both of the main classes of sparse linear solver algorithms. Finally, to show that our techniques are useful for end-to-end applications, we present Ōmeteōtl, an accelerator targeted at applications that use iterative solvers in their inner loop. Ōmeteōtl also shows that the techniques in this thesis generalize to sparse matrix computations beyond linear solvers.https://mit.zoom.us/j/98122373906 (no password)
TBD
January 22
Thesis Defense: Taming Data Movement Overheads in Latency-Critical Cloud Services
Nikita Lazarev
CSAIL
Add to Calendar
2025-01-22 15:00:00
2025-01-22 16:30:00
America/New_York
Thesis Defense: Taming Data Movement Overheads in Latency-Critical Cloud Services
Cloud providers are being urged to enhance the efficiency, performance, and reliability of datacenter infrastructures to support applications across many domains with diverse requirements for quality of service. Data movement is a significant source of overhead in today’s servers, and it is particularly critical for the recent emerging interactive and realtime cloud applications. In this thesis, I investigate and propose a set of novel approaches to mitigate the data movement overheads in general-purpose datacenters. This allows to establish a roadmap towards more efficient and reliable cloud services which are severely bottlenecked by data movement. In particular, I propose, implement, and evaluate three systems for the applications in (1) microservices, (2) serverelss, and (3) realtime cloud-native services on the example of virtualized radio access networks (vRAN), which are known to raise challenges to existing cloud infrastructures.First, we discuss Dagger – a system for mitigating the overheads of remote procedure calls in interactive cloud microservices. Dagger introduces a novel yet practical solution enabling fast and low-latency communication between distributed fine-granular application components. We then present Sabre – a practical and efficient system for mitigating the challenging overhead of cold start in serverless. Sabre relies on emerging tightly-coupled accelerators for compression and allows to dramatically reduce the latency of page movement in serverless microVMs without compromising the CPU cost. Finally, we build Slingshot – the first to the best of our knowledge infrastructure that enables fault tolerance in realtime cloud-native services such as vRAN. With Slingshot, we make substantial progress towards deploying reliable distributed systems working in realtime in the general purpose cloud by addressing the key challenges of fast state migration, realtime fault detection, and low-latency disaggregation.Thesis Committee: Christina Delimitrou (MIT), Zhiru Zhang (Cornell University), Mohammad Alizadeh (MIT)
TBD