Tuesday 16:35–17:05 in Track 3

Optimizing Deep Neural Network Layer Topology with Delve

Justin Shenk

Audience level:
Experienced

Description

Identifying the number of hidden units in a fully connected layer is considered a heuristically-guided craft. A PyTorch library, Delve, was developed that allows identifying the degree of over-parameterization of a layer, thus guiding architecture selection. The library compares the intrinsic dimensionality of the layer over training, providing the user with live feedback during training.

Abstract

Subscribe to Receive PyData Updates

Subscribe