Sure, in the end, everything is an RNN ;)

Mongillo's model indeed requires pre-formed attractors. However, that model opens up to other sub-threshold memory mechanisms that can maintain STM when LTMs are not present, yet.

So maybe if an RNN has memory that scales up with the size, it is a better model of activated LTM rather than of STM. No?

4-7 *new* items! If you learn to use a retrieval structure, you go pretty big - as all those crazy memorists do!

online WM of novel items require variable binding, and I doubt you can get it in RNNs.

Check this out for a cogsci perspective
pmc.ncbi.nlm.nih.gov/articles/PMC...

IMHO. Because it does not have "memory slots" ready for it, if it has memory slots, it lacks the machinery to convert them back into sensory representations. But you're the RNN expert, if you say it, I believe you ;)

Can it distinguish a shape it saw once from one it never saw? And despite visual distractors?

Maybe we do have five stable attractors in pfc and use them to re-cue sensory perceptions. But we need to be clear about active/silent interactions to move forward in this direction.

True. However, the attractor `iff` pers. activity is the case only if we consider the firing rate as the only variable at play in the brain. Any cognitive theory of WM requires the allocation of new items on the fly. Can we build attractors through single-shot learning? Or we repurpose the same?

True! However, we have learned about long-term memory formation and maintenance from snails. Divergent evolutionary paths can yet implement similar computational mechanisms. IMHO, recurrent activity is not as prominent in monkey's PFC as in fruitfly MB, but the comparison holds.

If you're still updating the starter pack, add me in ;)

I want to be the guy on the left. libera nos domine de AI

In the image I see a mid age person with a wise face and - probably - mastering a complex software Vs a youngster with a naive smile clicking buttons on an anonymous interface while the world burns

Then send a link to the blog ;)

And concluding with a beautiful thermodynamics of learning:

The correct learning rate depends on the noise of the training set (sigma eta)!

And reminding the great neuroscientist **and middle east pacifist** Daniel Amit!

The library strikes a balance between creating new models and composing them in fairly complex networks.

If you would like some specific feat and cannot achieve it, just contact me and we'll find a way together!

Where do they get all this money to infest our market anyway???

And yess! The students loved it :)

Thank you! I wanted to feel better this mornin 🤣