Sparsity-Probe: Analysis tool for Deep Learning Models

TL;DR

Sparsity-Probe is an analysis tool for deep learning models that evaluates intermediate layer representations using geometric features, aiding in understanding model depth contribution and identifying underperforming layers without extra test data.

Contribution

It introduces a novel sparsity-based probe grounded in machine learning and approximation theory to analyze deep neural network representations.

Findings

Enables measurement of layer contribution to overall performance.

Detects under-performing layers during or after training.

Does not require auxiliary test datasets.

Abstract

We propose a probe for the analysis of deep learning architectures that is based on machine learning and approximation theoretical principles. Given a deep learning architecture and a training set, during or after training, the Sparsity Probe allows to analyze the performance of intermediate layers by quantifying the geometrical features of representations of the training set. We show how the Sparsity Probe enables measuring the contribution of adding depth to a given architecture, to detect under-performing layers, etc., all this without any auxiliary test data set.