Azhir Mahmood  🚀 🤖 🧬 🧠 🪐

Counterpoint: I believe that all parameters, including channel densities etc are learned. Reason? Learning a channel is computationally identical to learning a synapse with constant input.

Sure, I don't disagree with changes in channel densities, spine growth, etc. But there are some aspects that probably can't be modified within a lifetime, e.g. molecular structure of a channel, or the gross architecture of the neuron. I.e. a distribution of constancy values

fully agree. Although, just to be contrarian, I bet that molecular configuration (but obviously not the amino acid sequence itself) and gross neuron architecture are both learned.

Do you mean within a lifetime or across generations? Obviously true for the latter, less clear for the former.

within a lifetime. Would not be contrarian otherwise I think ;)

I agree that we don't know given current methods if this is true or not. I'm inclined to believe that having too many dimensions of plasticity can impede learning so there is perhaps some virtue in having some more rigid parameters, esp those that allow for homeostasis. >

but isn't the main thing we learn from DL that giving more parameters tends to always help?

Only if the training set is able to scale up too right? Given a fixed amount of experience, it's not clear that more parameters is always better.