Embedding channel pruning within the CNN architecture design using a bi-level evolutionary approach.

Abstract

Remarkable advancements have been achieved in machine learning and computer vision through the utilization of deep neural networks. Among the most advantageous of these networks is the convolutional neural network (CNN). It has been used in pattern recognition, medical diagnosis, and signal processing, among other things. Actually, for these networks, the challenge of choosing hyperparameters is of utmost importance. The reason behind this is that as the number of layers rises, the search space grows exponentially. In addition, all known classical and evolutionary pruning algorithms require a trained or built architecture as input. During the design phase, none of them consider the process of pruning. In order to assess the effectiveness and efficiency of any architecture created, pruning of channels must be carried out before transmitting the dataset and computing classification errors. For instance, following pruning, an architecture of medium quality in terms of classification may transform into an architecture that is both highly light and accurate, and vice versa. There exist countless potential scenarios that could occur, which prompted us to develop a bi-level optimization approach for the entire process. The upper level involves generating the architecture while the lower level optimizes channel pruning. Evolutionary algorithms (EAs) have proven effective in bi-level optimization, leading us to adopt the co-evolutionary migration-based algorithm as a search engine for our bi-level architectural optimization problem in this research. Our proposed method, CNN-D-P (bi-level CNN design and pruning), was tested on the widely used image classification benchmark datasets, CIFAR-10, CIFAR-100 and ImageNet. Our suggested technique is validated by means of a set of comparison tests with regard to relevant state-of-the-art architectures.