What Do Compressed Deep Neural Networks Forget?

TL;DR

This paper investigates how model compression techniques like pruning and quantization affect the performance of deep neural networks on specific challenging data points, revealing that compression disproportionately impacts atypical and noisy images.

Contribution

It introduces the concept of Pruning Identified Exemplars (PIEs) and shows that compression impacts a small subset of data more severely, especially on long-tail, noisy, and atypical images.

Findings

Models with different sizes perform similarly overall but differ on PIEs.

Compression disproportionately affects long-tail, noisy, and atypical images.

PIEs are more challenging for both humans and algorithms to classify.

Abstract

Deep neural network pruning and quantization techniques have demonstrated it is possible to achieve high levels of compression with surprisingly little degradation to test set accuracy. However, this measure of performance conceals significant differences in how different classes and images are impacted by model compression techniques. We find that models with radically different numbers of weights have comparable top-line performance metrics but diverge considerably in behavior on a narrow subset of the dataset. This small subset of data points, which we term Pruning Identified Exemplars (PIEs) are systematically more impacted by the introduction of sparsity. Compression disproportionately impacts model performance on the underrepresented long-tail of the data distribution. PIEs over-index on atypical or noisy images that are far more challenging for both humans and algorithms to classify. Our work provides intuition into the role of capacity in deep neural networks and the trade-offs incurred by compression. An understanding of this disparate impact is critical given the widespread deployment of compressed models in the wild.