Dendrinet reveals synaptic clustering emerges from learning, boosting covariance discrimination

A new study by Ilenna Simone Jones and colleagues introduces Dendrinet, an artificial neural network that mimics the brain's dendritic structure with hierarchical segments and sparse conductance-based synapses. Trained on a Permuted-Covariance Classification (PCC) task—a problem that linear-nonlinear networks cannot solve—Dendrinet developed functional synapse clusters (FSCs) on its dendritic branches, mirroring experimental observations in cortical and hippocampal neurons. The researchers found that excitatory FSCs form only when both dendritic nonlinearities and synaptic structural plasticity are active; turning off the nonlinearities reduced excitatory FSCs and performance, while unexpectedly increasing inhibitory FSCs.

To test causal necessity, the team shuffled learned synaptic connectivity while keeping nonlinearities fixed. Performance dropped, and shuffling inhibitory synapses caused a larger decline than shuffling excitatory ones, revealing that inhibitory organization is more critical for solving the PCC task. This suggests that dendritic computation is not just a passive filter but an active computational unit that learns to cluster synapses for specific functions. The findings bridge experimental neuroscience and AI, showing how biological details like dendrites enable sophisticated pattern recognition that standard neural networks lack, and they open new avenues for bio-inspired machine learning architectures.