> And yet, artificial neural networks ARE an approximation of how biological neurons work. Only if you limit yourself to "sums of weighted inputs, sent through

> And yet, artificial neural networks ARE an approximation of how biological neurons work. Only if you limit yourself to "sums of weighted inputs, sent through a 1D activation function". However, the parent said "differentiable primitives": these days people have built networks that contain differentiable ray-tracers, differentiable physics simulations, etc. Those seem like crazy ideas if we limit ourselves to the "neural" analogy; but are quite natural for a "composition of differentiable primitives" approach.