Efficient Model Compression for Bayesian Neural Networks

TL;DR

This paper introduces a novel Bayesian-inspired model compression method for neural networks that uses posterior inclusion probabilities for pruning, resulting in models with better generalizability and efficiency.

Contribution

It presents a new Bayesian model selection-based pruning strategy for neural networks using spike-and-slab priors and variational inference.

Findings

Pruned models show improved generalization across benchmarks.

The method effectively reduces model complexity while maintaining performance.

Bayesian pruning enhances resistance to overfitting.

Abstract

Model Compression has drawn much attention within the deep learning community recently. Compressing a dense neural network offers many advantages including lower computation cost, deployability to devices of limited storage and memories, and resistance to adversarial attacks. This may be achieved via weight pruning or fully discarding certain input features. Here we demonstrate a novel strategy to emulate principles of Bayesian model selection in a deep learning setup. Given a fully connected Bayesian neural network with spike-and-slab priors trained via a variational algorithm, we obtain the posterior inclusion probability for every node that typically gets lost. We employ these probabilities for pruning and feature selection on a host of simulated and real-world benchmark data and find evidence of better generalizability of the pruned model in all our experiments.