What to Prune and What Not to Prune at Initialization

TL;DR

This paper introduces two novel initialization pruning methods, K-starts and dissipating gradients, that achieve high sparsity in neural networks without retraining or task-specific tuning.

Contribution

It presents new initialization pruning techniques that outperform existing methods in achieving sparsity while preserving network performance.

Findings

Combination of methods outperforms individual approaches and random dropout.

Methods do not require task-specific knowledge or retraining.

Achieves high sparsity with maintained performance.

Abstract

Post-training dropout based approaches achieve high sparsity and are well established means of deciphering problems relating to computational cost and overfitting in Neural Network architectures. Contrastingly, pruning at initialization is still far behind. Initialization pruning is more efficacious when it comes to scaling computation cost of the network. Furthermore, it handles overfitting just as well as post training dropout. In approbation of the above reasons, the paper presents two approaches to prune at initialization. The goal is to achieve higher sparsity while preserving performance. 1) K-starts, begins with k random p-sparse matrices at initialization. In the first couple of epochs the network then determines the "fittest" of these p-sparse matrices in an attempt to find the "lottery ticket" p-sparse network. The approach is adopted from how evolutionary algorithms find the best individual. Depending on the Neural Network architecture, fitness criteria can be based on magnitude of network weights, magnitude of gradient accumulation over an epoch or a combination of both. 2) Dissipating gradients approach, aims at eliminating weights that remain within a fraction of their initial value during the first couple of epochs. Removing weights in this manner despite their magnitude best preserves performance of the network. Contrarily, the approach also takes the most epochs to achieve higher sparsity. 3) Combination of dissipating gradients and kstarts outperforms either methods and random dropout consistently. The benefits of using the provided pertaining approaches are: 1) They do not require specific knowledge of the classification task, fixing of dropout threshold or regularization parameters 2) Retraining of the model is neither necessary nor affects the performance of the p-sparse network.