Streaming Networks: Increase Noise Robustness and Filter Diversity via Hard-wired and Input-induced Sparsity

Abstract

The CNNs have achieved a state-of-the-art performance in many applications. Recent studies illustrate that CNNs recognition accuracy drops drastically if images are noise corrupted. We focus on the problem of robust recognition accuracy of noise-corrupted images. We introduce a novel network architecture called Streaming Networks. Each stream is taking a certain intensity slice of the original image as an input, and stream parameters are trained independently. We use network capacity, hard-wired and input-induced sparsity as the dimensions for experiments. The results indicate that only the presence of both hard-wired and input-induces sparsity enables robust noisy image recognition. Streaming Nets is the only architecture which has both types of sparsity and exhibits higher robustness to noise. Finally, to illustrate increase in filter diversity we illustrate that a distribution of filter weights of the first conv layer gradually approaches uniform distribution as the degree of hard-wired and domain-induced sparsity and capacities increases.

Sparsifying and Down-scaling Networks to Increase Robustness to Distortions

“…In the paper [21], it was illustrated that input-induced sparsity (image intensity slices used as input into different streams) and hard-wired sparsity (parallel weight-decoupled streams) are both necessary for noise robustness to emerge in multi-stream architectures based on a simple conv net. STNet model based on the state-of-the-art models are build upon both input-induced sparsity and hard-wired sparsity.…”

“…STNet [20] has confirmed high robustness to noise, when tested on a random zero noise [21]. A random zero noise implies that values of random pixels are set to 0 for all color channels.…”