A MOEA/D approach to exploit the crucial structure of convolution kernels

Deep convolutional neural networks achieved remarkably performance on computer vision tasks. However, they are always over-parameterized and computationally intensive to deploy on computing platforms with limited hardware resources. In this paper, a MOEA/D based pruning approach is proposed. We exploit the crucial structure of convolution kernels using MOEA/D based pruning approach and find 88% computational redundancy in convolutional layers of LeNet-5 without any accuracy loss. Networks are encoded as chromosomes in continues or discrete spaces. A multi-objective evolutionary algorithm, MOEA/D, is employed to evolve multiple convolutional neural networks simultaneously subject to accuracy and computational cost as two conflicting objectives. Several experiments are conducted on three classical image classification tasks. Statistical results demonstrate that our proposed method is both effective and superior to other methods.