CDS Members at NeurIPS 2024

The Neural Information Processing Systems (NeurIPS) conference, taking place December 10–15 at the Vancouver Convention Center, will showcase cutting-edge research from around the globe. NeurIPS has long been one of the most influential venues in machine learning and artificial intelligence, known for featuring work that shapes the future of these fields.

This year, CDS faculty, researchers, and students will present research spanning an extraordinary range of topics, from theoretical advances in optimization and neural networks to practical applications in computer vision and natural language processing.

Here is the research they are presenting this year:

Rico Angell (Postdoc Researcher)

• Measuring Progress in Dictionary Learning for Language Model Interpretability with Board Game Models

Umang Bhatt (Faculty Fellow)

• Large Language Models Must Be Taught to Know What They Don’t Know

Sam Bowman (Associate Professor of Linguistics and Data Science)

• Many-shot Jailbreaking
• LLM Evaluators Recognize and Favor Their Own Generations

Joan Bruna (Professor of Computer Science and Data Science)

• Stochastic Optimal Control Matching
• Provable Posterior Sampling with Denoising Oracles via Tilted Transport

Lucius Bynum (PhD Student)

• Causal Dependence Plots

Angelica Chen (PhD Student)

• Preference Learning Algorithms Do Not Learn Preference Rankings

Kyunghyun Cho (Professor of Computer Science and Data Science)

• Iterative Reasoning Preference Optimization
• Jointly Modeling Inter- & Intra-Modality Dependencies for Multi-modal Learning
• Non-convolutional graph neural networks.
• Implicitly Guided Design with PropEn: Match your Data to Follow the Gradient
• Preference Learning Algorithms Do Not Learn Preference Rankings
• Multiple Physics Pretraining for Spatiotemporal Surrogate Models

Eunsol Choi (Assistant Professor of Computer Science and Data Science)

• SVFT: Parameter-Efficient Fine-Tuning with Singular Vectors

Yunzhen Feng (PhD Student)

• Model Collapse Demystified: The Case of Regression

Carlos Fernandez-Granda (Associate Professor of Mathematics and Data Science)

• Uncertainty-aware Fine-tuning of Segmentation Foundation Models

Siavash Golkar (Research Scientist)

• Multiple Physics Pretraining for Spatiotemporal Surrogate Models

Yuzhou Gu (Faculty Fellow)

• Log-concave Sampling from a Convex Body with a Barrier: a Robust and Unified Dikin Walk

Yanjun Han (Assistant Professor of Mathematics and Data Science)

• Assouad, Fano, and Le Cam with Interaction: A Unifying Lower Bound Framework and Characterization for Bandit Learnability
• Stochastic contextual bandits with graph feedback: from independence number to MAS number
• Online Estimation via Offline Estimation: An Information-Theoretic Framework

Shirley Ho (Senior Research Scientist)

• Multiple Physics Pretraining for Spatiotemporal Surrogate Models
• The Well: a Large-Scale Collection of Diverse Physics Simulations for Machine Learning
• The Multimodal Universe: Enabling Large-Scale Machine Learning with 100TBs of Astronomical Scientific Data

Chris Ick (PhD Student)

• Vocal Call Locator Benchmark (VCL’24) for localizing rodent vocalizations from multi-channel audio

Sanyam Kapoor (PhD Student)

• Large Language Models Must Be Taught to Know What They Don’t Know

Julia Kempe (Silver Professor of Computer Science, Mathematics, and Data Science)

• Mission Impossible: A Statistical Perspective on Jailbreaking LLMs
• Iteration Head: A Mechanistic Study of Chain-of-Thought
• The Price of Implicit Bias in Adversarially Robust Generalization
• Model Collapse Demystified: The Case of Regression

Yilun Kuang (PhD Student)

• Unlocking Tokens as Data Points for Generalization Bounds on Larger Language Models

Brenden Lake (Associate Professor of Psychology and Data Science)

• Beyond the Doors of Perception: Vision Transformers Represent Relations Between Objects

Yann LeCun (Professor of Computer Science, Neural Science, Data Science, and Electrical and Computer Engineering)

• Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs
• Fine-Tuning Large Vision-Language Models as Decision-Making Agents via Reinforcement Learning
• G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering

Sanae Lotfi (PhD Student)

• Unlocking Tokens as Data Points for Generalization Bounds on Larger Language Models

Qi Lei (Assistant Professor of Mathematics and Data Science)

• Stochastic Zeroth-Order Optimization under Strongly Convexity and Lipschitz Hessian: Minimax Sample Complexity
• Sketchy Moment Matching: Toward Fast and Provable Data Selection for Finetuning

Taro Makino (PhD Student)

• Jointly Modeling Inter- & Intra-Modality Dependencies for Multi-modal Learning

Aahlad Puli (Faculty Fellow)

• Explanations that reveal all through the definition of encoding
• Contrasting with Symile: Simple Model-Agnostic Representation Learning for Unlimited Modalities

Jonathan Niles-Weed (Associate Professor of Mathematics and Data Science)

• Learning Elastic Costs to Shape Monge Displacements
• Progressive Entropic Optimal Transport Solvers

Xiang Pan (PhD Student)

• Sketchy Moment Matching: Toward Fast and Provable Data Selection for Finetuning

Aram-Alexandre Pooladian (PhD Student)

• Learning Elastic Costs to Shape Monge Displacements
• Progressive Entropic Optimal Transport Solvers

Andres Potapczynski (PhD Student)

• Searching for Efficient Linear Layers over a Continuous Space of Structured Matrices

Rajesh Ranganath (Associate Professor of Computer Science and Data Science)

• Contrasting with Symile: Simple Model-Agnostic Representation Learning for Unlimited Modalities
• Preference Learning Algorithms Do Not Learn Preference Rankings
• Explanations that reveal all through the definition of encoding

Shauli Ravfogel (Faculty Fellow)

• On Affine Homotopy between Language Encoders

Mengye Ren (Assistant Professor of Computer Science and Data Science)

• Reawakening knowledge: Anticipatory recovery from catastrophic interference via structured training

Tim Rudner (Instructor)

• Pre-trained Text-to-Image Diffusion Models Are Versatile Representation Learners for Control

Cristina Savin (Associate Professor of Neural Science and Data Science)

• Learning predictable and robust neural representations by straightening image sequences

Jingtong Su (PhD Student)

• Mission Impossible: A Statistical Perspective on Jailbreaking LLMs

Ravid Shwartz-Ziv (Research Scientist)

• OpenDebateEvidence: A Massive-Scale Argument Mining and Summarization Dataset

Eero Simoncelli (Professor of Neural Science, Mathematics, Data Science, and Psychology)

• Learning predictable and robust neural representations by straightening image sequences
• Shaping the distribution of neural responses with interneurons in a recurrent circuit model
• Contrastive-Equivariant Self-Supervised Learning Improves Alignment with Primate Visual Area IT

Ilia Sucholutsky (Faculty Fellow)

• Learning Human-like Representations to Enable Learning Human Values

Yifei Sun (Adjunct Professor)

• G-Retriever: Retrieval-Augmented Generation for Textual Graph Understanding and Question Answering

Nikolaos Tsilivis (PhD Student)

• The Evolution of Statistical Induction Heads: In-Context Learning Markov Chains
• The Price of Implicit Bias in Adversarially Robust Generalization

Andrew Wilson (Professor of Computer Science and Data Science)

• Large Language Models Must Be Taught to Know What They Don’t Know
• Searching for Efficient Linear Layers over a Continuous Space of Structured Matrices
• Unlocking Tokens as Data Points for Generalization Bounds on Larger Language Models

Denny Wu (Faculty Fellow)

• Neural network learns low-dimensional polynomials with SGD near the information-theoretic limit
• Pretrained Transformer Efficiently Learns Low-Dimensional Target Functions In-Context

Lily Zhang (PhD Student)

• Preference Learning Algorithms Do Not Learn Preference Rankings

These contributions reflect the diversity of research interests at CDS, demonstrating how our members are advancing multiple frontiers in data science simultaneously. Their work ranges from foundational mathematical theory to practical applications in artificial intelligence, and from computational neuroscience to machine learning systems. The breadth and depth of these papers illustrate CDS’ commitment to pushing forward the boundaries of data science through rigorous, innovative research.