Yedi Zhang

I am a PhD candidate at the Gatsby Unit UCL, advised by Peter Latham and Andrew Saxe. I use theoretical approaches to study how neural networks with different architectures [1] learn, including attention-based [2], fully-connected [3], and multimodal [4] networks.
I recently spent a sunny winter at Stanford University working with Jay McClelland and Andrew Lampinen, supported by a Bogue fellowship. Previously, I spent summers working on 3D computer vision at DJI R&D and CUHK CSE.

Publication

	Saddle-to-Saddle Dynamics Explains A Simplicity Bias Across Neural Network Architectures Yedi Zhang, Andrew Saxe, Peter E. Latham ICLR 2026 webpage \| iclr \| openreview \| arxiv \| emoji \| talk
	Training Dynamics of In-Context Learning in Linear Attention Yedi Zhang, Aaditya K. Singh, Peter E. Latham, Andrew Saxe ICML 2025 (Spotlight) pmlr \| openreview \| arxiv \| code \| talk
	Understanding Unimodal Bias in Multimodal Deep Linear Networks Yedi Zhang, Peter E. Latham, Andrew Saxe ICML 2024 webpage \| pmlr \| arxiv \| code

Conference proceeding

Saddle-to-saddle dynamics explains a simplicity bias across neural network architectures
Yedi Zhang, Andrew Saxe, Peter E. Latham
International Conference on Learning Representations (ICLR), 2026
Training dynamics of in-context learning in linear attention
Yedi Zhang, Aaditya K. Singh, Peter E. Latham*, Andrew Saxe*
International Conference on Machine Learning (ICML), Spotlight, 2025
Understanding unimodal bias in multimodal deep linear networks
Yedi Zhang, Peter E. Latham, Andrew Saxe
International Conference on Machine Learning (ICML), 2024

Journal article

When are bias-free ReLU networks effectively linear networks?
Yedi Zhang, Andrew Saxe, Peter E. Latham
Transactions on Machine Learning Research (TMLR), 2025

Preprint

To use or not to use muon: how simplicity bias in optimizers matters
Sara Dragutinović, Yedi Zhang, Rajesh Ranganath
2026

Non-archival conference

Transitivity can redeem the reversal curse in attention models
Yedi Zhang, Andrew Lampinen*, James McClelland*
Annual Conference on Cognitive Computational Neuroscience (CCN), 2026
Optimal learning rate scaling depends on data in deep scalar linear networks
Yedi Zhang, Peter E. Latham, Leena Chennuru Vankadara, Andrew Saxe
ICLR Workshop on Scientific Methods for Understanding Deep Learning, 2026

Bio

Publication

Blog

Fun