(A) model study on the long-range horizontal connectivity organizing small-world networks in visual cortex

Abstract

Both our brain and deep neural networks (DNNs) successfully perform object recognition, but the visual pathway in the brain consists of fewer hierarchical layers than a DNN, presumably due to the restricted volume of the brain. What is the distinct strategy of the brain as compared to the artificial network that enables cost-efficient visual processing under this physical constraint, and how can this be implemented in artificial neural networks? Here, I suggest that the cortical long-range connections (LRCs) observed in various mammalian species can organize balanced local and global connections in the network, and they enable object recognition even in shallow neural networks. Using simulations of a model hierarchical network with convergent feedforward connections and LRCs from tree shrew data, I investigated the functional roles of LRCs for object recognition. First, I observed that the addition of LRCs to a shallow feedforward network significantly enhances the CIFAR-10 image classification performance so that it is comparable to that of a much deeper network. Second, through network pruning with gradient-based optimization, I confirmed that LRCs could spontaneously emerge by minimizing the total connection length while maximizing classification performance. Although most of the initial connections were pruned during training, a certain portion of the very long connections survived until the end of training. Ablation of the surviving LRCs led to a significant reduction in classification performance, which implies that LRCs are crucial for image classification. Lastly, I found that a combination of sparse LRCs and dense local connections organizes small-world type connectivity in a model network, and this enables consistent recognition of various types of objects. Furthermore, I found that performance enhancement by LRCs is strongly correlated with the small-worldness of a network, and this can possibly explain the optimal ratio of LRCs in the visual cortex. Taken together, I propose that long-range horizontal connectivity might be a key architecture of the visual cortex to implement parsimonious object recognition under the physical constraints of the brain.