Poster 3

Derek Montanez

Texas State University

Exploring the Tradeoffs Between Compression and Accuracy in Deep Neural Networks

Abstract: Machine learning, specifically neural networks, are at the heart of many important everyday jobs such as image classification, and natural language processing tasks. Advances in today’s neural networks, have led to very impressive results in varying tasks, with the impressive results comes the disadvantage of very large neural network models, with parameters in the range of hundred millions to billions. There has been a lot of ongoing research to be able to provide solutions to the problem of compressing large neural network models, to perform as well as the original model or even better. The goal currently is to be able to compress deep neural networks to be able to deploy and use on edge devices, failure to deploy large models is inevitable due to memory and computational constraints on edge devices. In this paper, we will explore the trade-offs between compression and accuracy in neural networks using a neural network compression method ’Pruning Filters for Efficient ConvNets’. Neural network compression methods are used to reduce the memory space and computational costs of neural networks, and these methods include but are not limited to pruning, quantization, and knowledge distillation. The tests will be conducted using the the MNIST Fashion data-set, as well as the CIFAR10 data-set, both of which are image classification data-sets. The model SCNNB is a convolutional neural network that will be used for conducting tests.