Competitive learning, temporal memory learning and spatial pooler learning

Competitive learning, a form of Hebbian learning, has been around since the 80s. It’s been some time I’ve not followed Numenta’s work, but, from my understanding, the way the spatial pooler and the temporal memory learn is quite similar to e.g. self-organising maps (which learn in a competitive way).

I’m interested in a comparison between competitive learning and the way the SP and TM learn. What are the differences and similarities? They both learn in an unsupervised way, but I am interested in more details. (I will eventually answer to this question, if nobody is able to do it, but only once I have some more time.)

This might be a bit of a stretch to call standard HTM a SOM system.

In the classic SOM the winning cell pulls “next best neighbors” along with them with some scaled version of the training reward. In HTM it’s winner-take-all and the losers get nothing.

This subject is near and dear to my heart as I have been trying to work out how to make the Cortical IO semantic maps in some sort of online system. I have not seen how to do that yet but I keep thinking about it.

3 Likes