Hacker News Reader

Comment by ithkuil | Hacker News Reader

When a neural network "learns" something, it ends up with knowledge encoded in the network parameters. If that's not memory, we're using terminology in a different way.

View on HN

I don't disagree that this is a possible way of viewing memory. I'm am simply restating the argument that I've heard the best I know how. That usually comes with a "memory is an illusion" caveat. If you want to argue that that is memory instead of a typical "hard drive" sort of situation then sure, okay, but this is not what most people think of when they think "memory". They believe it is something in which individual memories could perhaps be extracted, implanted, etc. I personally believe it's a valid theory, however it lacks experimental evidence and has yet to account for the fact that we can recall individual "training data". To account for this I do think it would require fundamental changes in how we view "computation". If you believe that artificial neural network structures can account for the brains real neuron processing I suggest checking out the book I mentioned, not because it cant be true but that there are still big unknowns left to discover before we know for sure if this is correct.[0]

[0]Memory and the Computational Brain, C. R. Gallistel and Adam Philip King

You can read the main thesis here https://onlinelibrary.wiley.com/doi/pdf/10.1002/978144431049...

I have no idea; this is not my field. Just pointing out it's quite possible people are just talking through each other.

Human language is like that, full of ambiguity, of implicit baggage.

If you feel "memory" means a recording of an event, or a photo book, then faced with the discovery that brains actually don't work like that you may want to state "neurons don't store memories!".

Reply on news.ycombinator.com