Striving for Simplicity: The All Convolutional Net

The paper questions the standard design of convolutional neural networks for object recognition, which typically alternate convolution and max-pooling layers before a few fully connected layers. Re-evaluating the state of the art for object recognition from small images, the authors test whether each pipeline component is truly necessary, and find that max-pooling can simply be replaced by a convolutional layer with increased stride without any loss in accuracy on several image recognition benchmarks.

Building on this finding, they propose a simplified architecture consisting solely of convolutional layers, which achieves competitive or state-of-the-art performance on CIFAR-10, CIFAR-100, and ImageNet. To analyze the network, they also introduce a new variant of the deconvolution approach for visualizing features learned by CNNs, applicable to a broader range of network structures than existing approaches, reinforcing a trend toward simpler, more uniform network designs.