DL Fundamentals Lab
Texas A&M University
Deep Learning Fundamentals Lab

Tomer
Galanti

Assistant Professor · Computer Science & Engineering
Texas A&M University · galanti@tamu.edu · Office: 325 PETR
Publications Research Group Google Scholar Full Bio
Pier Beneventano
Research Focus

Building a rigorous science of modern AI

Tomer Galanti is an Assistant Professor in the Department of Computer Science and Engineering at Texas A&M University. His research develops the mathematical foundations of modern AI — from understanding why pre-trained representations generalize and transfer, to building provably capable systems from LLM agents.

Prior to joining Texas A&M, he was a postdoctoral associate at MIT's Center for Brains, Minds & Machines, working with Tomaso Poggio. He received his Ph.D. from Tel Aviv University, advised by Lior Wolf, and interned as a Research Scientist at Google DeepMind in 2021.

Prior to joining Texas A&M, he was a postdoctoral associate at MIT's Center for Brains, Minds & Machines, working with Tomaso Poggio. He received his Ph.D. from Tel Aviv University, advised by Lior Wolf, and interned as a Research Scientist at Google DeepMind in 2021.

Selected Research Contributions
All publications →
01
Distributed Speculative Inference (DSI)
A framework that parallelizes speculative decoding across multiple devices, making LLM inference provably and empirically faster — without sacrificing output quality.
ICLR 2025 LLM Inference Efficiency
02
Formation of Representations in Neural Networks
A theoretical account of how structured representations emerge during training, explaining the geometry of learned features in deep classifiers.
ICLR 2025 Spotlight Representation Theory
03
DisCO: Reinforcing Large Reasoning Models
Discriminative Constrained Optimization for RLHF training of reasoning-capable LLMs, achieving stronger performance with more stable training dynamics.
NeurIPS 2025 LLM Reasoning RLHF
04
On the Role of Neural Collapse in Transfer Learning
Shows that neural collapse — the geometric alignment of final-layer features to class means — is the key mechanism behind the success of pretrained representations for transfer.
ICLR 2022 Neural Collapse Transfer Learning
05
On the Modularity of Hypernetworks
A rigorous theoretical analysis of hypernetworks, establishing when and why parameter-generating networks induce modular structure — with implications for continual learning and multi-task settings.
NeurIPS 2020 Oral Hypernetworks
06
On the Power of Decision Trees in Auto-Regressive Language Modeling
Proves that decision tree ensembles are expressive enough to implement next-token prediction, providing a new theoretical lens on the computational substrate of LLMs.
NeurIPS 2024 LLM Theory Expressivity
07
Reverse Engineering Self-Supervised Learning
Decomposes SSL objectives to reveal the implicit supervision signal, bridging the gap between contrastive learning and classical supervised approaches. Co-authored with Yann LeCun.
NeurIPS 2023 Self-Supervised Learning
08
SGD and Weight Decay Secretly Minimize Rank
Proves that standard training with SGD and weight decay implicitly drives neural network weight matrices toward low rank — explaining compressibility and generalization from first principles.
CPAL 2025 Implicit Regularization Generalization
Teaching
Special Topics in Recent Developments in Deep Learning & LLMs Texas A&M University · Fall 2025
Introduction to Machine Learning Texas A&M University · Spring 2025
Special Topics in Recent Developments in Deep Learning & LLMs Texas A&M University · Fall 2024
Statistical Learning Theory and its Applications Massachusetts Institute of Technology · Fall 2023
Statistical Learning Theory and its Applications Massachusetts Institute of Technology · Fall 2022
Deep Convolutional Neural Networks Tel Aviv University · Spring 2020
Deep Convolutional Neural Networks Tel Aviv University · Spring 2019