Learning Sparse Filters in Deep Convolutional Neural Networks with a $$l_1/l_2$$ Pseudo-Norm

While deep neural networks (DNNs) have proven to be efficient for numerous tasks, they come at a high memory and computation cost, thus making them impractical on resource-limited devices. However, these are known contain large number of parameters. Recent research has shown that their structure can more compact without compromising performance.In this paper, we present sparsity-inducing regularization term based the ratio \(l_1/l_2\) pseudo-norm defined filter coefficients. By defining appropriately different kernels, removing irrelevant filters, kernels in each layer drastically reduced leading very Deep Convolutional Neural Networks (DCNN) structures. Unlike existing methods, our approach does not require an iterative retraining process and, using term, directly produces sparse model during training process. Furthermore, is also much easier simpler implement than methods. Experimental results MNIST CIFAR-10 show significantly reduces filters classical models such as LeNet VGG while reaching same or even better accuracy baseline models. Moreover, trade-off between sparsity compared other loss terms l1 l2 norm well SSL [], NISP [] GAL methods shows outperforming them.