I’m very excited avout the latest developments and research around sensorimotor integration.
We have discussed with a friend about the use of HTM Theory / Nupic for anomaly detection and I was wondering if sensorimotor integration can also be used in such cases meaning is SMI just that (sensor/motor) or is it a more generic feature?
HTM Neuron, SDRs are very generic terms but SMI is specific to objects/locations, or is it? I suppose “context” is an abstraction of the feature which is higher level ? I think Jeff spoke about this somewhere.
The term SMI makes a lot of sense for basic human intelligent behavior but aren’t those features of the cortex used for higher/abstract context (or action) integration?
The way I see it the neuron is closest to neuroscience but if higher level features can be abstracted away from neuroscience, like is the case with SDRs compared to SMI, it would be better for HTM Theory so it’s more easily understood and utilized.
This could be a question for Nupic and handled in the actual implementation but I think HTM Theory needs to stay as generic as possible.
I’ve been looking into HTM Theory on a very high level so I could be missing a lot of context in which case even a quick pointer to more information would be enough of an answer.
@dwrrehman made some interesting observations in this thread which are relevant to what you are touching on here.
The concept of location (which is a foundation of SMI) could be looked at in a couple of ways. One is the obvious correlation to a physical position (cells activating in a grid pattern which depicts one’s position in a room, for example). The other is to encode semantics in otherwise randomized cell activations, whereby semantically similar locations share active SDR bits.
From there, the concept of “location” can be taken to higher levels of abstraction, such as the sense of knowing “where” a particular memory is in your brain when you “reach out” to play it back. This type of abstraction can be applied to other concepts as well, such as the “motor commands” that you use when you are humming a melody in your head, etc.
It really is quite generic. The SMI circuit is just an example of what someone might create with it.
Thanks for the reference to @dwrrehman’s observation!
It seems though grid cells are mostly studied in how they help form metric spatial representation and maybe my suggestion that SMI might be used in a more abstract fashion is deviating from the facts. Your example with the location of a memory is also interesting so at least some of my thoughts are in a good direction. It seems the cortex is trying to do all kinds of tricks to reuse its features in ingenious ways.
Just found out this video with Jeff and Matt (4m 57s):
ha, id say that’s even an understatement! according to my theory, at least.
it says that there are actually only two fundamental things that the cortex uses in order to do everything it needs to do: an inference layer, and a pooling layer.
…which, also according to my theory, are really the same thing, except for one little difference.