A possible neural-network representation of the Kernel memory.

<p>Architecture (a) corresponds to the algorithm of learning and recall as described in the text. In (b) we use an approximation to maximize the capacity and reduce the number of neural units. Middle layer with synaptic matrix is eliminated, synaptic matrices of the resulting two layers are identical up to transposition. Therefore we have distinct connections ( is the number of stored memories and is the input dimension). We can choose a “primary” layer to store the connections, the other one will mirror them. Effective memory capacity of this architecture is .