📢Abstract submissions is 16th March AOE for the Annual Conference on the Mathematics of Neuroscience and AI (May 27-30th; National Opera Theatre, Split, Croatia)

A truly outstanding line-up of keynotes, session chairs, and invited speakers. We'd love you to come along too!

www.neuromonster.org

Now how’s this all related to cognitive maps and hippocampus? We show that this slot algorithm is exactly the same maths as TEM just with some equations reshaped and rearranged. Frontal and temporal cognitive maps get unified! (8/9)

Slots don’t just have to be flat, they can be hierarchical. We show that in a hierarchical task, RNNs learn hierarchical slots! A prediction for experimentalists to test 😀. (7/9)

Slots also work for structured behavioural tasks too (Basu et al & El-Gaby et al). A continuous manifold of slots account for PFC neurons including ‘progress’ and ‘structured memory buffer’ neurons. These ‘behavioural’ tasks are really just working memory tasks in disguise! (6/9)

Armed with this theory, we can reinterpret PFC representations as slots! Not just in simple sequence repetition tasks (Xie et al), but also cue-dependent recall tasks (Panichello et al). Here the cue acts like a velocity signal to move memories between slots 😀. (5/9)

And slots are exactly what randomly initialised RNNs/SSMs actually learn! And for many many task structures and sizes (not just the one shown in the figure here) (4/9)

Our new slot algorithm works for any recurrent network (PFC / RNN / SSM) for any structured sequence memory task. This is a big generalisation of prior mechanisms (e.g. Botvinick or Ganguli) that could only repeat the exact sequence it was presented. (3/9)

What’s the new mechanism? Different memories get stored in different neural subspaces–‘slots’. Importantly, slots are connected to each other (mirroring task connectivity!) so that memories get passed between slots allowing the right memory to be readout at the right time! (2/9)