A Global Approach for Goal-Driven Pruning of Object Recognition Networks

Abstract

Pruning methods for neural network models are important for devices with performance and storage problems. Recently, unlike traditional pruning methods, The Goal Driven Pruning method has been proposed. This approach, inspired by the attention mechanism in humans, is based on decreasing the sensitivity to the features of distractors in the environment. For this purpose, in this method, pruning is performed not only in the middle layers, but also in the output layers for the task irrelevant classes. In this study, we present Global Goal-driven Pruning, which, unlike Goal-driven Pruning, prunes by evaluating the model as a whole, instead of layer-based pruning. The effectiveness of the proposed model has been demonstrated by the tests.