A Selective Multi-Unit Group Filter-Based Flexible Fast Light-CNN Training

Grouped convolution is a method extensively used in the design of Convolutional Neural Network (CNN) architecture. Its independent divided grouped filter facilitates the learning and provide possibility of learning various type of features. This reduces the number of parameters and computation costs, making it particularly useful for design lightweight CNN models, which is applicated for mobile devices. While several techniques such as group-based network pruning have been proposed to further improve its efficiency, these methods often require complex additional algorithms that could be challenging to implement on hardware. In this paper, we propose Multi-Unit Group Filter Block (MUGB), which is an efficient and simple grouped convolution method without the requirement of additional algorithms to improve the performance. MUGB can provides various combinations through selectively activate multiple grouped filters in response to the learning difficulty of the dataset or the target learning speed of the operation. This fast method of training that specialized in limited resources, and dataset, based on small-scale CNN models, could maintain high performance compare to other methods of similar compression ratio. At the compression rate of 81.2%, our proposed model achieves an accuracy of 1.5% higher than a similarly-sized CNN model, and 3.5% higher than a pruned model, while maintaining the performance of the uncompressed standard CNN model.