Dendrite assemblies as the core cortical computation module for continual motor learning

The cortex has the amazing capacity to continuously learn through experience while retaining past memories.

But how does the cortical network implement this continual learning while avoiding interference and catastrophic overwriting of prior events?

While cell assemblies with simple point neurons are thought to serve as the basic learning and storage units, this model poses major challenges in dynamic environments and lacks experimental support.

Relying on strong preliminary results, I here propose a radically different view of learning and storage in the cortex—the dendrite assembly hypothesis—where the relevant memory units are the “hidden layer” of dendritic branches.

Namely, each neuron operates as a small network, with different dendrite assemblies representing different tasks and driving the soma.

The dendrite assembly model augments the cell assembly model, potentially alleviating problems of interference, sparsity and capacity.

We will test the dendrite assembly hypothesis in the mouse motor cortex, where learning is perpetual and coding is dense.

This will entail determining dendritic and somatic representations during continual learning, thus deciphering the core learning units of the network (Aim1), the pathways (Aim2) and structural plasticity (Aim3) that enable dendrite assembly formation and learning; and the consequences of the dendrite assembly model for the pathogenesis of Parkinson’s disease (Aim4).

We will record from somas, dendrites and spines of pyramidal tract neurons at single-cell and population levels with unprecedented spatiotemporal resolution, using state-of-the-art in-vivo imaging, a novel behavioral design, and an analysis platform we developed.

Our results are expected to transform our view of how cortical neurons represent multiple motor memories in the healthy and Parkinsonian brain, open avenues for developing novel treatment modalities for Parkinson’s disease and inspire new artificial intelligence network architectures.