HLC How to generate behavior with HTM?

Ok. Thanks. I have seen that diagram often enough. I guess I should have known that.

Hmm, that’s true.

I understand your point in regards to understanding the human behavior, in this case you need a body too. The insight I think Jeff had was to assume that the neocortex is an intelligent open system i.e. it is a sub-system of a larger system. In a radical simplification, from the “perspective” of the neocortex the environment is the rest of the brain. In HTM it is clear the “theory” is intended to have the HTM generate behavior - this is present in the very early work of Jeff I think.

Your ideas are not compatible with HTM in that regard but if the distinction is kept clear then I think it is fine. When you pick one piece of HTM as “right” I would be very wary, as it is not intended to mimic the brain - it is wildly different from the brain (to being with it is digital and the brain is analog).


From my current understanding of the role of dendrites, I would state it more like this:

Each dendrite is sampling from the input space (an SDR). That is to say they are each monitoring a specific subset of input neurons. They are a coincidence detector in that they are waiting for the moment when some significant fraction of those watched neurons are active at the same time. When the number of active sampled neurons are above the specified threshold, then the dendrite depolarizes it’s parent neuron, which puts it in either a predictive state (distal/axial) or active state (proximal).

So the dendrites are not forming SDR’s themselves, but are rather sampling from a subset of nodes that are part of the input (or contextual) SDR. The dendrite itself does not know which specific SDR (input pattern) is being detected. However, one of our modeling assumptions is that SDRs with sufficient overlap (share some subset of active neurons) are semantically or contextually similar. Therefore, the desired behavior is for the coincidence detectors (dendrites) to consistently respond to the simultaneous activation of its sampled neurons as if it is detecting a semantically and/or contextually significant feature in the input.


just some general observations:
think of the Neocortex as a giant memory of sequences of actions with associated sensory inputs.
What would a “memory access” look like to extract a known sequence to achieve a goal?
It can not be the same action/sensory input, that just tells you what you already know.
You need some form of “guided” internal activation to access and build a sequence.
Without details, that is part of what the cerebellum does. Thats why the cerebellum is involved in reasoning and memory retrieval.
Since HTM is very biologically inspired you should study up on that :wink:
My working hypothesis is that grid cells are involved in that.
Some thing like so:
You have a goal and then you build or retrieve from memory a sequence of actions to achieve it.
This sequence will consist of repeating sub sequences
these repeating sub sequences of actions define a set of metrics on the way to your goal
and that way you know where you are …