Avatar

Noam Razin

Postdoctoral Fellow

Princeton Language and Intelligence, Princeton University

 

I am a Postdoctoral Fellow at Princeton Language and Intelligence. Previously, I completed my PhD in Computer Science at Tel Aviv University, where I was fortunate to be advised by Nadav Cohen.

My research focuses on the foundations of modern machine learning. I aim to develop theories that shed light on how neural networks work, as well as bring forth principled methods for improving their efficiency, reliability, and performance.

My work is supported in part by a Zuckerman Postdoctoral Scholarship. During my PhD, I interned at Apple Machine Learning Research and the Microsoft Recommendations Team, and received the Apple Scholars in AI/ML and the Tel Aviv University Center for AI & Data Science fellowships.

Email: noamrazin (at) princeton.edu

🗞 News

  • Oct 24: New paper proving that the implicit bias of state space models (SSMs) can be poisoned with clean labels.

  • Oct 24: When aligning language models via Direct Preference Optimization (DPO), past work observed that the probability of preferred responses often decreases. So where does the probability go? 🧐 In a new paper, we characterize the causes for this counter-intuitive phenomenon, show that it can lead to surprising failures in alignment, and provide preventative guidelines.

  • Oct 24: Honored to receive the Zuckerman Postdoctoral Scholarship and the Israeli Council for Higher Education Postdoctoral Scholarship.

  • Sep 24: Joined Princeton Language and Intelligence as a Postdoctoral Fellow.

  • Aug 24: New lecture notes on the theory (and surprising practical applications) of linear neural networks.

  • May 24: Implicit Bias of Policy Gradient in Linear Quadratic Control: Extrapolation to Unseen Initial States accepted to ICML 2024.

  • Jan 24: Two papers accepted to ICLR 2024: one identifying a vanishing gradients problem when using reinforcement learning to finetune language models and another analyzing length generalization of Transformers.

  • Sep 23: Two papers accepted to NeurIPS 2023: one on the ability of graph neural networks to model interactions and another on what makes data suitable for locally connected neural networks.

  • Sep 23: Interned at Apple Machine Learning Research.

📄 Publications

* denotes equal contribution

The Implicit Bias of Structured State Space Models Can Be Poisoned With Clean Labels
Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization
Understanding Deep Learning via Notions of Rank
Implicit Bias of Policy Gradient in Linear Quadratic Control: Extrapolation to Unseen Initial States
Vanishing Gradients in Reinforcement Finetuning of Language Models
What Algorithms Can Transformers Learn? A Study in Length Generalization
Lecture Notes on Linear Neural Networks: A Tale of Optimization and Generalization in Deep Learning
What Makes Data Suitable for a Locally Connected Neural Network? A Necessary and Sufficient Condition Based on Quantum Entanglement
On the Ability of Graph Neural Networks to Model Interactions Between Vertices
Implicit Regularization in Hierarchical Tensor Factorization and Deep Convolutional Neural Networks
Implicit Regularization in Tensor Factorization
Implicit Regularization in Deep Learning May Not Be Explainable by Norms
RecoBERT: A Catalog Language Model for Text-Based Recommendations
Scalable Attentive Sentence-Pair Modeling via Distilled Sentence Embedding

💬 Selected Talks

  • Understanding and Overcoming Failures of Language Model Finetuning
    The Flatiron Institute Machine Learning Seminar, November 2024
    Slides

  • Analyses of Policy Gradient for Language Model Finetuning and Optimal Control
    MPI MiS + UCLA Math Machine Learning Seminar, March 2024
    Video Slides

  • Two Analyses of Modern Deep Learning: Graph Neural Networks and Language Model Finetuning
    Princeton Alg-ML Seminar, December 2023
    Slides

  • On the Ability of Graph Neural Networks to Model Interactions Between Vertices
    Learning on Graphs and Geometry Reading Group, January 2023
    Video Slides

  • Generalization in Deep Learning Through the Lens of Implicit Rank Lowering
    ICTP Youth in High-Dimensions: Recent Progress in Machine Learning, High-Dimensional Statistics and Inference, June 2022
    Video Slides

  • Generalization in Deep Learning Through the Lens of Implicit Rank Lowering
    MPI MiS + UCLA Math Machine Learning Seminar, May 2022
    Slides

  • Implicit Regularization in Tensor Factorization
    The Hebrew University Machine Learning Club, Jerusalem, Israel, June 2021
    Video Slides

  • Implicit Regularization in Deep Learning May Not Be Explainable by Norms
    Tel Aviv University Machine Learning Seminar, Tel Aviv, Israel, May 2020
    Slides

👨‍🏫 Teaching

  • Teaching Assistant for Foundations of Deep Learning (course #0368-3080), Tel Aviv University, 2021 to 2023