Hybrid Genetic Based Algorithm for CNN Ultra Compression
Abdelrhman Mohamed Abotaleb, Ahmed Tarek Elwakil, Mayada Hadhoud · 2019
The increasing deployment of the Convolutional Neural Networks (CNNs) in computer vision algorithms is because of the high accuracy of CNNs in image recognition jobs which makes CNNs of dominant use, but the huge number of parameters inside the CNN puts hurdles against the ease of deployment of the CNN in embedded systems, IoT chips and even mobile devices. Huge number of parameters means excessive computations and storage requirements which makes it infeasible to use in hardware chips with limited resources. In this paper a new hybrid algorithm for CNNs compression is proposed. This new algorithm combines pruning, quantization and compression techniques which are used to reduce the huge size of different CNNs while maintaining a similar accuracy percentage. The proposed algorithm is to combine the use of genetic algorithms (GA) to conduct the selection criteria of pruning of convolutional layers only and applying conventional pruning to the fully connected layers, Finally applying the quantization and compression on the resulted neural network weights.