Reducing architectural complexity of convolutional neural networks via channel pruning

Abstract

The architectural complexity of a neural network is reduced by selectively pruning channels. A cost metric for a convolution layer is determined. The cost metric indicates a resource cost per channel for the channels of the layer. Training the neural network includes, for channels of the layer, updating a channel-scaling coefficient based on the cost metric. The channel-scaling coefficient linearly scales the output of the channel. A constant channel is identified based on the channel-scaling coefficients. The neural network is updated by pruning the constant channel. Model weights are updated via a stochastic gradient descent of a training loss function evaluated on training data. The channel-scaling coefficients are updated via an iterative-thresholding algorithm that penalizes a batch normalization loss function based on the cost metric for the layer and a norm of the channel-scaling coefficients. When the layer is batch normalized, the channel-scaling coefficients are batch normalization scaling coefficients.