We designed a visual-auditory association task to demonstrate the applicability of this ‘assembly calculus’ framework in a real-world example. Our model can correctly classify letters and numbers in a downstream concept area, even if only part of the sensory information is presented. 9/12

We next investigated how the formed assemblies can be combined across areas. We focus on two assembly operations – projections and associations. In short, these operations can be learned across areas – importantly without any decay (forgetting) of previously learned structures. 8/12

In a recurrent network with 400 multi-compartment neurons, switching one inhibitory context “on” allows for learning of stable dendrite-specific assemblies at disinhibited dendritic compartments. 7/12

Assuming that inhibitory (sub-) populations encode “context” signals controlling distinct dendritic compartments, inhibitory neurons can flexibly gate plasticity on or off at each specific dendritic compartment. 6/12

We use a multi-compartment neuron model with a spiking soma in which synapses at the dendrites undergo voltage-dependent synaptic plasticity. Here, the balance of excitatory and inhibitory inputs at each dendrite determines the sign and amount of plasticity. 5/12

We propose a biological plausible model, combining nonlinear dendrites and inhibitory context-dependent gating to enable flexible learning without forgetting. Our model bridges scales from dendritic properties to assembly learning in recurrent circuits to multi-area assembly computations. 4/12