| Abstrakt: |
Nowadays, convolutional neural networks (CNNs) have achieved tremendous performance in many machine learning areas. However, using a large number of parameters leads to the redundancy problem, which negatively impacts the performance of CNNs. Indeed, many kernels are redundant and can be taken off from the network without much loss of performance. In this paper, we propose a new optimization model for localizing and removing the redundancy in CNN. In fact, Unlike numerous existing methods where they only reduce the redundancy, our proposal also allows to localize the distribution of redundancy in CNNs. The suggested model consists of two stages: in the first one, a dataset is used to train a specific CNN generating a learned CNN with optimal parameters. These later are combined with a decision L 1 -sparsity optimization model for detecting and reducing the unwanted kernels. At the end, the evolutionary genetic algorithm is adapted to solve the proposed model generating finally an optimal CNN with prior information about the redundancy distribution. The performance of our approach has been shown and demonstrated by several experiments. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink