Events

Hide Filters Show Filters
Hide Filters Show Filters

Event Type

Impact Area

Research Area

Seminar Series

Reset Filters

November 09, 2024

No events scheduled

November 12, 2024

Automatically Optimizing Distributed Protocols

David Chu
University of California Berkeley
1:30P
- 2:30P

Location

32-G449 (Patil/Kiva)
Add to Calendar 2024-11-12 13:30:00 2024-11-12 14:30:00 America/New_York Automatically Optimizing Distributed Protocols Bio: David Chu is a PhD student at Berkeley advised by Natacha Crooks and Joe Hellerstein. He works across databases and distributed systems, using query optimization and database theory to optimize distributed systems, and using those optimized systems to push the state of the art in databases.Abstract: Distributed protocols like Paxos and PBFT often form the bottleneck of distributed systems and databases. To solve this, researchers have created a large suite of consensus and BFT variants, but this process is error-prone and many variants contain bugs. We present a new, orthogonal approach: Instead of creating new protocols from scratch, we introduce a set of correct-by-construction rewrites that can be used by a compiler to automatically optimize any distributed protocol. 32-G449 (Patil/Kiva)

MIT ML Seminar: Swaroop Mishra

Swaroop Mishra
Google DeepMind
2:00P
- 3:00P

Location

32-D463 (Star)
Add to Calendar 2024-11-12 14:00:00 2024-11-12 15:00:00 America/New_York MIT ML Seminar: Swaroop Mishra Title: Instruction Following and Reasoning in Large Language ModelsBio: Swaroop Mishra is a Senior Research Scientist at Google DeepMind (formerly Google Brain), where he works on Gemini reasoning. He and his team recently built a system that received a Silver Medal at the International Mathematical Olympiad (IMO) 2024. His main research contributions are on instruction-tuning and reasoning methods, including natural instructions, super-natural instructions, reframing, question-decomposition, math via programs, help me think, and instruction-bias (EACL 2023 Outstanding Paper Award). He is a co-organizer of the MATH-AI workshop at NeurIPS 2022 and NeurIPS 2024. His work on "Natural Instructions" has recently received the AI2 "Lasting Impact Paper Award." 32-D463 (Star)

Panel Discussion: Open Questions in Theory of Learning

Moderator: Tomaso Poggio - Panel: Ila Fiete, Philip Isola, Eran Malach, Haim Sompolinsky
4:00P
- 5:30P

Location

46-3002
Singleton Auditorium
Add to Calendar 2024-11-12 16:00:00 2024-11-12 17:30:00 America/New_York Panel Discussion: Open Questions in Theory of Learning Abstract: In a society that is confronting the new age of AI in which LLMs begin to display aspects of human intelligence, understanding the fundamental theory of deep learning and applying it to real systems is a compelling and urgent need.This panel will introduce some new simple foundational results in the theory of supervised learning. It will also discuss open problems in the theory of learning, including problems specific to neuroscience.Panelists:Ila Fiete - Professor of Brain and Cognitive Sciences, MITHaim Sompilinski - Professor of Molecular and Cellular Biology and of Physics, Harvard UniversityEran Malach - Research fellow, Kempner Institute at Harvard UniversityPhilip Isola - Associate Professor, EECS at MIT 46-3002

November 13, 2024

Augmenting AI Methods With Data Management Systems

Michael Cafarella
CSAIL MIT
4:00P
- 5:00P

Location

Kiva 32-G449
Add to Calendar 2024-11-13 16:00:00 2024-11-13 17:00:00 America/New_York Augmenting AI Methods With Data Management Systems Title: Augmenting AI Methods With Data Management SystemsAbstract: Over the past decade-plus, AI advances have yielded wonder after wonder. However, AI methods often require an expensive scaffolding of data curation and data engineering before they can be fully exploited. This talk will describe two separate projects to build data systems around AI methods and thereby make them more practical, scalable, and widely applicable. The first data system applies causal reasoning methods to the problem of debugging distributed systems using log data. We show that the system can find root causes of problems 1.6-18x more accurately than competing methods, while reducing the amount of human effort by a factor of 2.5-16.8 when compared to more conventional curation practices in causal inference. The second data system enables declarative optimization for foundation model programs. We show that when compared with conventional AI program implementations across several benchmark tasks, even the single-threaded optimized programs can reduce runtime by 50-83%, can reduce model costs by 40-86%, and can retain output quality that is roughly similar to the original.Bio:Michael Cafarella is a Principal Research Scientist in the MIT Computer Science and Artificial Intelligence Lab. His research interests include databases, information extraction, data integration, and applying data-intensive methods to economics. He has published extensively in venues such as SIGMOD, VLDB, and elsewhere. His academic awards include the NSF CAREER award, the Sloan Research Fellowship, and the VLDB Ten-Year Best Paper award. In addition to his academic work, he costarted the widely-used Hadoop open-source project, and cofounded Lattice Data, a research-based startup that was later acquired by Apple. From 2009 to 2019 he was a professor of Computer Science and Engineering at the University of Michigan. Kiva 32-G449

November 14, 2024

Cardinality Estimation using Lp norms on Degree Sequences

Dan Olteanu
University of Zurich
1:00P
- 2:00P

Location

Star - 32-D463
Add to Calendar 2024-11-14 13:00:00 2024-11-14 14:00:00 America/New_York Cardinality Estimation using Lp norms on Degree Sequences AbstractCardinality estimation is the problem of estimating the size of the output of a query, without actually evaluating the query. The cardinality estimator is a critical piece of any query optimizer, and is often the main culprit when the optimizer chooses a poor plan.In this talk, I will introduce LpBound, a pessimistic cardinality estimator that computes a guaranteed upper bound on the size of the query output. Besides common statistics used for cardinality estimation, such as the sizes of the input relations, the domain sizes for various columns, histograms, and the most common key values in join columns, LpBound also uses Lp norms on the degree sequences of the join columns. The computed cardinality bound is an optimal solution of a linear program whose constraints encode such statistics and Shannon inequalities.I will also report on an on-going experimental evaluation of LpBound for a variety of queries from the JOB, STATS, and subgraph matching benchmarks: acyclic and cyclic queries, full and group-by queries, and queries with and without range and equality predicates. LpBound takes time and uses memory very close to the traditional estimators used by PostgreSQL, DuckDB, and a commercial database system. Yet the range of LpBound overestimates is much smaller than the range of the estimates of the traditional estimators, which can both underestimate and overestimate by orders of magnitude. When fed the LpBound cardinalities, PostgreSQL can take less query evaluation time than when using its own estimates.This is joint work with Mahmoud Abo Khamis (RelationalAI), Chris Mayer (U Zurich), Dan Suciu (U Washington), and Haozhe Zhang (U Zurich).BioDan Olteanu is a professor at the University of Zurich, where he leads the Data Systems and Theory group (https://www.ifi.uzh.ch/en/dast.html), and a computer scientist at RelationalAI (https://relational.ai), where he works on incremental view maintenance, cardinality estimation, in-database machine learning, and data science projects.

November 15, 2024

Seminar Dot-Product Proofs and Their Applications

Ron Rothblum, Succinct and Technion
10:30A
- 12:00P

Location

32-G882, Hewlett Room
Add to Calendar 2024-11-15 10:30:00 2024-11-15 12:00:00 America/New_York Seminar Dot-Product Proofs and Their Applications We introduce a new type of probabilistic proof-system, called a dot-product proof (DPP). In a DPP, the input statement x and the proof \pi are vectors over a finite field F, and the proof is verified by making a single dot-product query to the concatenation of x and \pi. We will discuss constructions of DPPs and their applications, including:1. Exponential hardness of approximation for MAXLIN.2. Extremely succinct argument-systems. Joint work with Nir Bitansky, Prahladh Harsha, Yuval Ishai and David Wu 32-G882, Hewlett Room

November 19, 2024

Thesis Defense, Lirui Wang. Robot Fleet Learning from Heterogeneous Data

Lirui Wang
CSAIL MIT
9:00A
- 11:00A

Location

Room 32-D463 (STAR)
Add to Calendar 2024-11-19 9:00:00 2024-11-19 11:00:00 America/New_York Thesis Defense, Lirui Wang. Robot Fleet Learning from Heterogeneous Data TALK: Thesis Defense: Robot Fleet Learning from Heterogeneous DataThesis Defense: Robot Fleet Learning from Heterogeneous DataSpeaker: Lirui WangIn-person Location: 32-D463 (Star)Zoom Link: https://mit.zoom.us/j/99424131532Abstract:One of the key roadblocks for training generalist robotic models today is heterogeneity. Previous robot learning methods often collect data to train with one specific embodiment for one task, which is expensive and prone to overfitting. Similar to humans, robots and embodied agents inherently have to deal with heterogeneous inputs and outputs due to the nature of the perception-action loops across diverse environments. The data format and distributions collected from these systems and used for training them are varied in different modalities such as color, depth, tactile, and proprioceptive information, and/or collected in different domains such as simulation, real robots, and human videos. Moreover, fleets of robots and machines ingest massive amounts of streaming data generated by interacting with their environments in a distributed fashion, and teams of robots shall co-acquire diverse skills through their experiences in varied settings. The core idea behind my research, fleet learning, is to embrace the heterogeneous nature of robot learning to develop efficient and general algorithms. In this thesis, I will present a few angles toward tackling such challenging problems and application domains, ranging from tokenizing data, aligning representations, and merging policies, to composing skills. We develop insights and theories, often from linear settings, for how fleet learning can lead to more principled and effective use of robotic data and propose algorithmic progress, often through alignments, toward building generalist robotic foundation models. Empirically, we show advanced robotic manipulation capabilities by leveraging data from multimodal sensory inputs and multiple domains. In addition to outperforming several previous state-of-the-art across simulation and real-world benchmarks, we develop intelligent systems for robotic applications such as package handling in warehouses as well as dexterous tool-use tasks that have applications such as manufacturing, logistics, and household robots.

November 21, 2024

Geometric Algebra Planes: Convex Implicit Neural Volumes

Sara Fridovich-Keil
Stanford University

Part Of

11:00A
- 12:00P

Location

32-D507
Add to Calendar 2024-11-21 11:00:00 2024-11-21 12:00:00 America/New_York Geometric Algebra Planes: Convex Implicit Neural Volumes Volume parameterizations abound in recent literature, from the classic voxel grid to the implicit neural representation and everything in between. While implicit representations have shown impressive capacity and better memory efficiency compared to voxel grids, to date they require training via nonconvex optimization. This nonconvex training process can be slow to converge and sensitive to initialization and hyperparameter choices that affect the final converged result. We introduce a family of models, GA-Planes, that is the first class of implicit neural volume representations that can be trained by convex optimization. GA-Planes models include any combination of features stored in tensor basis elements, fol- lowed by a neural feature decoder. They generalize many existing representations and can be adapted for convex, semiconvex, or nonconvex training as needed for different inverse problems. In the 2D setting with a linear feature decoder, we prove that GA-Planes is equivalent to a low-rank plus low-resolution matrix factorization; we show that this approximation outperforms the classic low-rank plus sparse decomposition for fitting a natural image. In 3D, we demonstrate GA-Planes’ competitive performance in terms of expressiveness, model size, and optimizability across three volume fitting tasks: radiance field reconstruction, 3D segmentation, and video segmentation. 32-D507

December 01, 2024

Cancelled

Test event

This event has been cancelled

12:00P
- 1:00P

December 03, 2024

MIT ML Seminar: Eric Vanden-Eijnden

Eric Vanden Eijnden
NYU
2:00P
- 3:00P

Location

32-G449 (Patil/Kiva)
Add to Calendar 2024-12-03 14:00:00 2024-12-03 15:00:00 America/New_York MIT ML Seminar: Eric Vanden-Eijnden Title: Generative modeling with flows and diffusionsAbstract: Generative models based on dynamical transport have recently led to significant advances in unsupervised learning. At mathematical level, these models are primarily designed around the construction of a map between two probability distributions that transform samples from the first into samples from the second. While these methods were first introduced in the context of image generation, they have found a wide range of applications, including in scientific computing where they offer interesting ways to reconsider complex problems once thought intractable because of the curse of dimensionality. In this talk, I will discuss the mathematical underpinning of generative models based on flows and diffusions, and show how a better understanding of their inner workings can help improve their design. These results indicate how to structure the transport to best reach complex target distributions while maintaining computational efficiency, both at learning and sampling stages. I will also discuss applications of generative AI in scientific computing, in particular in the context of Monte Carlo sampling, with applications to the statistical mechanics and Bayesian inference, as well as probabilistic forecasting, with application to fluid dynamics and atmosphere/ocean science.Bio: Eric Vanden-Eijnden is a Professor at NYU Courant working on ML foundations and applications to scientific computing. 32-G449 (Patil/Kiva)

Quest | CBMM Seminar Series: Kelsey Allen

Kelsey Allen
DeepMind
4:00P
- 5:30P

Location

46-3002
Singleton Auditorium
Add to Calendar 2024-12-03 16:00:00 2024-12-03 17:30:00 America/New_York Quest | CBMM Seminar Series: Kelsey Allen Kelsey Allen is a Senior Research Scientist at Deepmind. Her work surrounds cognitive science, machine learning, and robotics, with a focus on tool use and tool creation. 46-3002

December 10, 2024

February 19, 2025