Year: 2024
Type: preprint
Abstract: Drawing motivation from the manifold hypothesis, which posits that most high-dimensional data lies on or near low-dimensional manifolds, we apply manifold learning to the space of neural networks. We learn manifolds where datapoints are neural networks by introducing a distance between the hidden layer representations of the neural networks. These distances are then fed to the non-linear dimensionality reduction algorithm PHATE to create a manifold of neural networks. We characterize this manifold using features of the representation, including class separation, hierarchical cluster structure, spectral entropy, and topological structure. Our analysis reveals that high-performing networks cluster together in the manifold, displaying consistent embedding patterns across all these features. Finally, we demonstrate the utility of this approach for guiding hyperparameter optimization and neural architecture search by sampling from the manifold.
Source: arXiv (Cornell University)
Authors: Abel, Elliott, Steindl, Andrew J., Mazioud, Selma, Schueler, Ellie, Ogundipe, Folu, Zhang, Ellen, Grinspan, Yvan, Reimann, Kristof, Crevasse, Peyton, Bhaskar, Dhananjay, Viswanath, Siddharth, Zhang, Yanlei, Rudner, Tim G. J., Adelstein, Ian, Krishnaswamy, Smita
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Awards: Multiscale data geometric networks for learning representations and dynamics of biological systems, Collaborative Research: III: Medium: Contextualized and Multimodal Foundation Models for Graph Data in Scientific Discovery, REU Site: Summer Undergraduate Math Research at Yale, G6671297155, CAREER: Deep representation learning for exploration and inference in biomedical data
OpenAlex ID: W4404574161