Neuromorphic computation from algorithms to physical systems

I want to resist the usual shortcut of picking a neuromorphic chip first and then asking what it can do. The order I prefer is: primitives → dynamics → substrate.

Primitives before silicon

Before any hardware mapping, I am trying to pin down the computational primitives:

• sparse, event-driven representations rather than dense tensors;
• local plasticity rules of the form $\Delta w_{ij} = \eta\, x_i\, (y_j - \bar y_j)$, where updates depend only on locally available signals;
• memory as a dynamical state, not a separate addressable store.

Only once those primitives are clear does the question "which substrate?" become well-posed — because now there is a specification a substrate must satisfy, rather than a chip looking for a workload.

A running list of open questions lives at the end of this note and will keep expanding.