Blur and Knowledge from Falsehood: Neural Network Science and Neurophysiology Meets Epistemology

Abstract

Convolutional neural networks (CNNs) trained on clear images fail catastrophically with degraded or blurry imagery. New results by Jang and Tong, and Pramod, Katti and Arun show visual object recognition is optimized by introducing peripheral blur. Optimizing recognition of objects this way empirically supports the significance of there being a hundred times less photoreceptors dedicated for peripheral vision than in the retina. These results refute a longstanding epistemic slogan: Knowledge of truths arises only from knowledge of truths. Blur-trained CNNs and humans recognize things in blurry, degraded and noisy environments—a dog, a radiator—that clear-image-trained CNNs don’t. Blurring is misinformation about what is seen, so the human perceptual system recognizes objects by processes that start from falsehood. Peripheral blur—misinformation about what is seen—is essential to perceptual knowledge.