Brain-inspired Robust Vision using Convolutional Neural Networks with Feedback
Primates have a remarkable ability to correctly classify images even in the presence of significant noise and degradation. In contrast, even the state-of-art CNNs are extremely vulnerable to imperceptible level of noise. Many neuroscience studies have suggested that robustness in human vision arises from the interaction between the feedforward signals from bottom-up pathways of the visual cortex and the feedback signals from the top-down pathways. Motivated by this, we propose a new neuro-inspired model, namely Convolutional Neural Networks with Feedback (CNN-F). CNN-F augments CNN with a feedback generative network that shares the same set of weights along with an additional set of latent variables. CNN-F combines bottom-up and top-down inference through approximate loopy belief propagation to obtain the MAP-estimates of the latent variables. We show that CNN-F’s iterative inference allows for disentanglement of latent variables across layers. We validate the advantages of CNN-F over the baseline CNN in multiple ways. Our experimental results suggest that the CNN-F is more robust to image degradation such as pixel noise, occlusion, and blur than the corresponding CNN. Furthermore, we show that the CNN-F is capable of restoring original images from the degraded ones with high reconstruction accuracy while introducing negligible artifacts.
PDF Abstract
