Emergence of V1 connectivity pattern and Hebbian rule in a performance-optimized artificial neural network

The connectivity pattern and function of the recurrent connections in the primary visual cortex (V1) have been studied for a long time. But the underlying mechanism remains elusive. We hypothesize that the recurrent connectivity is a result of performance optimization in recognizing images. To test this idea, we added recurrent connections within the first convolutional layer in a standard convolutional neural network, mimicking the recurrent connections in the V1, then trained the network for image classification using the back-propagation algorithm. We found that the trained connectivity pattern was similar to those discovered in biological experiments. According to their connectivity, the neurons were categorized into simple and complex neurons. The recurrent synaptic weight between two simple neurons is determined by the inner product of their receptive fields, which is consistent with the Hebbian rule. Functionally, the recurrent connections linearly amplify the feedforward inputs to simple neurons and determine the properties of complex neurons. The agreement between the model results and biological findings suggests that it is possible to use deep learning to further our understanding of the connectome.