Modelling hippocampal activity using temporal memory

So I recently came across this paper from Vicarious where they use their novel HMM model (which they dub “cloned HMM”) to model the processes of the hippocampus in constructing cognitive maps. The model that they use in question is a standard Hidden Markov Model, with the only difference being that the transition matrix is sparse and deterministic.

The current state of the CLA is “complete” enough (I can’t think of a better word) to perform almost any of the same sequence tasks as the current standards for sequence modelling (LSTM, HMM etc.). As such, I find that it may be possible in theory to construct cognitive maps in the same way as the aforementioned paper, only replacing the HMM model with HTM (with necessary modifications of course). I find that the benefits of HTM (online learning, continuous learning) would certainly improve, and further biologically constrain, the model of the hippocampus that the researchers proposed.

What are your thoughts?


Hey @ayaz-amin, cool!

Sounds worth testing at least!

From a glance it seems a data type they use are simple categories (‘A’,‘B’,‘E’ etc), which make it easy to test since there’s an encoder for it.

My first questions would be:

  • What’s the scoring function? Like next-character prediction or something more complex?
  • Is the data shared anywhere?

If we can access the data and know how to evaluate the HTM outputs I think it’s quite doable.