[60654]

There appear to be two very interesting results:

1. We can observe how the state machine gets generated, first just a jumble of locations in a hub and spokes topology (no correlations), then some correlations start happening pairwise, making a kind of a beads on a string topology, and then finally the mental model snaps marvelously to two completely separate paths that meet at ends. It's amazing to see these mental models get formed in vivo out of initial unstructured perceptions.

2. In addition to standard HMM modeling, authors find that a "biologically plausible recurrent neural network (RNN) trained using Hebbian learning" can mimic some of this (but not exactly). But more interestingly, they find that LSTMs or transformers cannot. Which makes sense structurally, but it's a good reminder for those who believe the anthropomorphic hype that transformers have memory or other such (they don't :) ).

The scanning is indeed very intrusive, though.

[calf]

Couldn't memoryless neural networks still possibly learn the Next-State function of a Finite-State machine? Depending on the training algorithm. Especially if the eventual usage of such networks is to be called over and over again to generate the next token; conceptually this to me seems analogous to the process of finitely unrolling a while loop or a computer pipeline.