Suzanna Parkinson

Suzanna Parkinson

Computational and Applied Math PhD Candidate

ReLU Neural Networks with Linear Layers are Biased Towards Single- and Multi-Index Models

Published in arXiv, 2023

Connects the representation cost of neural networks with 1 ReLU layer and many linear layers to the spectrum of the expected gradient outer product matrix (EGOP), showing that this architecture is biased towards single- and multi-index models.

Joint work with Greg Ongie and Rebecca Willett.

https://arxiv.org/abs/2305.15598