Synaptic clustering emerges from learning and supports covariance discrimination
Functional synapse clusters (FSCs) are synapses with correlated presynaptic activity that are colocalized on the same neuronal dendritic branch. FSCs have been observed after learning in cortical and hippocampal pyramidal neurons. However, previous efforts to ablate FSCs by pharmacologically blocking dendritic nonlinearities to establish causal necessity may have confounded effects. Therefore, whether FSCs are causally necessary for computation is unknown. Here, we attempt to isolate FSCs from this potential confounder in silico. We train Dendrinet, an artificial neural network architecture with hierarchical dendritic segments and sparse conductance-based synapses, on a Permuted-Covariance Classification (PCC) task. This task cannot be solved by single-layer linear-nonlinear artificial neural networks. We find that neurons with dendrites can be trained to solve the task and develop excitatory and inhibitory FSCs if both dendritic nonlinearities and synaptic structural plasticity are active. Turning off dendritic nonlinearities reduces excitatory FSCs, which replicates experimental findings, and reduces performance while unexpectedly increasing inhibitory FSCs. Furthermore, shuffling learned synaptic connectivity while keeping the nonlinearities fixed reduces performance. This shows sensitivity to learned connectivity, but the shuffle does not change only FSCs. Shuffling inhibitory synapse properties reduces performance more than the corresponding excitatory shuffle, showing higher sensitivity to inhibitory organization. This work suggests that dendritic compartmentalization and learned synaptic organization can support computation of covariance structure.