TL;DR: CNNs with biologically-inspired lateral connections learned in an unsupervised manner are more robust to noisy inputs.
Keywords: lateral connections, convolutional neural networks, extra-classical receptive fields, mouse V1, supervised and unsupervised learning
Abstract: In the visual system, neurons respond to a patch of the input known as their classical receptive field (RF), and can be modulated by stimuli in the surround. These interactions are often mediated by lateral connections, giving rise to extra-classical RFs. We use supervised learning via backpropagation to learn feedforward connections, combined with an unsupervised learning rule to learn lateral connections between units within a convolutional neural network. These connections allow each unit to integrate information from its surround, generating extra-classical receptive fields for the units in our new proposed model (CNNEx). We demonstrate that these connections make the network more robust and achieve better performance on noisy versions of the MNIST and CIFAR-10 datasets. Although the image statistics of MNIST and CIFAR-10 differ greatly, the same unsupervised learning rule generalized to both datasets. Our framework can potentially be applied to networks trained on other tasks, with the learned lateral connections aiding the computations implemented by feedforward connections when the input is unreliable.