Maxmin convolutional neural networks for image classification
This work addresses a fundamental issue in CNN design for computer vision researchers, but it appears incremental as it builds on existing architectures with a specific modification.
The paper tackles the problem of how information and invariance are encoded in deep convolutional neural networks (CNNs) for image classification by proposing a MaxMin strategy that modifies the convolutional block to exploit both positive and negative high scores in convolution maps, resulting in improved performance over standard CNNs on MNIST and CIFAR-10 datasets.
Convolutional neural networks (CNN) are widely used in computer vision, especially in image classification. However, the way in which information and invariance properties are encoded through in deep CNN architectures is still an open question. In this paper, we propose to modify the standard convo- lutional block of CNN in order to transfer more information layer after layer while keeping some invariance within the net- work. Our main idea is to exploit both positive and negative high scores obtained in the convolution maps. This behav- ior is obtained by modifying the traditional activation func- tion step before pooling. We are doubling the maps with spe- cific activations functions, called MaxMin strategy, in order to achieve our pipeline. Extensive experiments on two classical datasets, MNIST and CIFAR-10, show that our deep MaxMin convolutional net outperforms standard CNN.