It’s been 2 month ago I have been posted here about mixing HTM and RL
After all I decided to use HTM into RL in order to solve credit assignment problem encounter in Reinforcement Learning theory. (1st idea)
Now I am told by my supervisor to seek for the second idea which should be related to HTM. Hence I wanted to know the current drawbacks of HTM and how can I improve it (seeking for the second idea).
NB: it’s left with 7 months to publish a paper and write my master s thesis.
Hey guys, know that I am not in your league. I’m a uni drop-out, so I’m probably not well-placed to give you academic advice.
However, in my very humble opinion, the new holy grail of artificial intelligence research, and something that is highly related to HTM, are grid cells.
There are two ways to address this:
How do grid cells produce a grid: what is the underlying principle that creates the behavior of neurons to fire in a grid pattern. And how can it be functionally simulated.
How to implement grid cells into HTM: how to use grid cell input and combine it in a meaningful manner with sensorial input and/or stored SDR input to produce a model of perceived reality.
Anyone who advances our understanding in this area deserves a Turing award.
RL has been using tile coding to discretize continuous state spaces for quite a while, especially before we got neural networks working as good function approximators. See page 217 of the book. If you are heading into the direction @Falco suggested, that might be a good place to start.
My suggestion would be to try using temporal memory as the model in a model-based algorithm.
Thanks,
sorry to disturb you once again but based on your suggestion and after reading around i have decided to deal with HTM mostly concentrate on Grid Cells
but there is few things i would like to clear:
1.How grid cells are selected in the neocortex?
it seems that grid cell are randomly selected in the neocortex.
2.How can grid cells be modify to improve HTM framework?
