Evaluating the Lottery Ticket Hypothesis to Sparsify Neural Networks for Time Series Classification

Abstract

Reducing the complexity of deep learning models is a challenging task in many machine learning pipelines. In particular for increasingly complex data spaces, the question of how to mitigate storage efforts for large machine learning models becomes of crucial importance. The recently proposed Lottery Ticket Hypothesis is one promising approach in order to decrease the size of a neural network without losing its expressiveness. While the Lottery Ticket Hypothesis has been shown to outperform other pruning methods in the field of image classification, it has not yet been extensively investigated in the domain of time series. In this paper, we thus investigate this hypothesis for the task of time series classification and empirically show that different deep learning architectures can be compressed by large factors without sacrificing expressiveness.