Lightweight Neural Networks

TL;DR

Lightweight Neural Networks use binary weights (+1, -1, or zero) to achieve near state-of-the-art accuracy with minimal storage, fast computation, and effective pruning, tested on multiple datasets.

Contribution

This paper introduces a neural network architecture with predominantly binary weights, emphasizing discretization and pruning for efficiency and comparable accuracy.

Findings

Achieved similar accuracy to conventional networks on MNIST and credit datasets.

Used networks with up to 16 hidden layers and 4.4 million weights.

Demonstrated effective weight discretization and pruning techniques.

Abstract

Most of the weights in a Lightweight Neural Network have a value of zero, while the remaining ones are either +1 or -1. These universal approximators require approximately 1.1 bits/weight of storage, posses a quick forward pass and achieve classification accuracies similar to conventional continuous-weight networks. Their training regimen focuses on error reduction initially, but later emphasizes discretization of weights. They ignore insignificant inputs, remove unnecessary weights, and drop unneeded hidden neurons. We have successfully tested them on the MNIST, credit card fraud, and credit card defaults data sets using networks having 2 to 16 hidden layers and up to 4.4 million weights.