It’s as the title suggests. Is integrating HTM with such a heuristic doable, like AlphaZero / MuZero? At the moment, the CLA is not mathematically grounded so there is not much that can be done, though theoretically it should work by training the TM via a modified Hebbian learning mechanism that supports supervision.
1 Like
Good point. It will be awesome is someone can pull it off with combination of the 1k brains theory. It might boost HTM to have NN level accuracy/performance for those tasks.
Tho I think CLA (or at least the SP and TM) is in fact very mathematically grounded. They are basically STDP but with large time steps. And training with supervision has been possible for a long time now, We don’t talk much about it because a single Spatial Pooler is not a very strong learner and the method is biologically impossible.
1 Like