[Paper Review] What Do Compressed Deep Neural Networks Forget?

Sara Hooker, Aaron Courville+3 more|arXiv (Cornell University)|Nov 13, 2019

Advanced Neural Network Applications86 references85 citations

TL;DR

The paper audits how pruning and quantization affect classifier performance beyond top-line accuracy, revealing a small set of instances (PIEs) that are disproportionately impacted by compression and methods differing in their disparate impact.

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.

Motivation & Objective

• Motivate and quantify how network compression (pruning and quantization) impacts generalization beyond overall accuracy.
• Identify whether certain classes or exemplars are disproportionately affected by compression.
• Develop a formal framework to audit disparate harm in compressed models and quantify trade-offs.
• Examine how compression interacts with distribution shifts and robustness to corruptions.

Proposed method

• Define a framework to measure class-level impact by comparing relative class recall under compression to baseline accuracy.
• Train populations of K models per compression setting to create distributions of class-wise accuracy for statistical testing (Welch’s t-test).
• Compute a relative recall deviation per class to control for overall accuracy changes.
• Introduce Pruning Identified Exemplars (PIEs) as images where predictions of pruned and non-pruned models disagree on the modal label.
• Evaluate pruning and three post-training quantization schemes across CIFAR-10, CelebA, and ImageNet with multiple architectures and levels of sparsity.

Experimental results

Research questions

• RQ1 Do compression methods impact some classes more than others, beyond changes in overall accuracy?
• RQ2 What are PIEs and what characteristics do they exhibit compared to non-PIEs?
• RQ3 How does compression affect robustness to distribution shifts and corruptions?
• RQ4 Is quantization inherently less disparately harmful than pruning across tasks?

Key findings

• Top-line accuracy can mask substantial non-uniform impact across classes under compression.
• A small subset of classes shows statistically significant recall degradation under pruning, increasing with sparsity (e.g., 170 significant classes at 50% sparsity; 372 at 70%).
• PIEs are disproportionately challenging for both artificial and human classification, with PIEs becoming more prevalent at higher sparsity (10.27% at 90% sparsity for ImageNet).
• Pruned models exhibit greater sensitivity to natural adversarial images and corruptions, amplified at higher sparsity levels.
• Quantization tends to produce less disparate harm than pruning; the most aggressive int8 fixed-point quantization affects more classes than moderate pruning, but still less than high-sparsity pruning in relative terms.
• Non-uniform impact varies with task complexity; CIFAR-10 shows fewer impacted classes than ImageNet, suggesting overparameterization and task difficulty modulate disparate effects.