I know HTM is quite far from having a working motor and attentional models, but is there some kind of computational model of working memory in the current literature that is at least semi-compatible with HTM?
I don’t really know anything (i’m just an aspiring comp neuro + htm builder guy) BUT, I think you should look into Long Short Term Memory Networks, Recurrent Neural Networks, Differentiable Neural Computers AND perhaps most helpful is Florian Fiebigs PhD thesis titled “Active Memory Processing on Multiple Time-scales in Simulated Cortical Networks with Hebbian Plasticity”. He discusses a novel model of working memory. Here’s the link for that: http://www.diva-portal.org/smash/record.jsf?pid=diva2%3A1263428&dswid=-4906. Florian Fiebigs visited/worked with Numenta and HTM so maybe his model is perhaps a good candidate as far as compatibility goes.
HTM memory works by pattern completion. Many neural models work the same way and there is a considerable amount of theory to back up the utility of this computing model. Very few (none?) also handle temporal memory
Each map in the sheet of columns has (potentially) a vast number of overlapping learned temporal/spatial patterns. When the perception is registered the saved pattern that most closely resembles the sensed pattern is recalled. There is some parsing to distribute these memories & recall over the hierarchy
This acts as a filter AND directs attention to novelty. This goes directly to your attention point.
What is passed to the subcortical structures is a basket of features.
The initiation of action is a task delegated to these subcortical structures. The subcortex projects patterns to the frontal lobe to be unfolded by pattern completion in the varous maps until it reaches the central sulcus. This goes to your motor point.
By some relatively complex interactions, the expressive part of the cortex (frontal lobe) learns patterns that are further refined by the cerebellum to make the motor commands more flexible; this is still pattern completion.
Note that the recalled patterns are not tagged - there is no A=>B transformation in the HTM model.
The HTM pattern recall and parsing do extract central tendencies and present a crisp pattern to the subcortical structures. The fact that some pattern is NOT recognized is sufficient to distinguish novelty; the bursting behavior signals this to the thalamus, and through that, to other subcortical structures.
If i’m not wrong, he is asking about the working memory (aka short term memory).
Resonating with perception is short term memory. (This is the pattern completion mentioned above) These traces decay rapidly, depending on the number of maps and subcortical structures that form your current conscious perception of “now;” a few seconds to a few minutes.
The circulation from highest level of the hierarchy (temporal lobe), through subcortex, out to the forbrain, and projections from the forbrain back to the sensor hierarchy form the basic loop of consciousness. The reverberating/recirculating structure holds the contents of “now.” Each pass though this loop reinforces/refreshes the contents of consciousness.
The key to longer term memory is the registration in the EC/HC complex. Until the conscious perception is captured by the EC/HC it is the effervescent now.
I (and I know others here) have been working on different models… at least for trying to create an agent, here’s an (imperfect, incomplete, but at least a stab) idea for how to create a working memory… its three levels of HTM with some manual hebbian-like rules (slightly modified CLA). Too busy at the moment creating a platform and trying to get it off the ground, and as such, haven’t tested it. But I do think it, or something like it, would be a viable approach to try.
Anyhow, for what it’s worth:
But EC has the largest plasticity, right? BTW, HTM homogeneous/constant plasticity is a problem.
That is a tricky point to answer.
The bulk of the cortex learns by “simple” Hebbian process - relatively slow. The fact that patients like HM are capable of learning skills like mirror writing without any memory of having done so reflects that the episodic memory is what is missing when the EC/HC is compromised.
Due to the distributed and interlocking nature of the representation, the cortex capacity in enormous. To understand this; there are only a small number of letters in an alphabet but due to combination and the rules of grammar they can represent a large number of objects and relationships - the concept is much the same.
The EC/HC has a much faster learning rate but relatively small capacity. Also, this learning rate is modulated by other subcortical structures, notably, the amygdala. The copy held in the EC/HC is like an index key in a database - replay during sleep reactivates the weak new cortical memories formed during the day, strengthening them.
The key to conscious episodic recall is a well formed pattern from the EC/HC to trigger pattern completion.
The functional divide between the cortex and EC/HC is seamless - it allows you to recall episodes throughout the day. The cortex has a exponential decay rate that drops a memory below consciousness in minutes without refreshing. The EC/HC can access these weak traces anytime during the day, and hammers on them with spindle waves to make them permanent at night.
As far as an agent interacting with the world, watch videos of patients with severely compromised EC/HC. They talk, they reason, they perceive and understand, they move and manipulate space with ease. And after a few minutes they completely forget anything that happened; once it leaves the “here and now” it is gone forever.
I think you are right. Nevertheless, let me answer with some crazy idea
I think that the cortex is be very reluctant to modify any synapse (unless is clearly needed, i.e. novel data). Beyond global modulatory mechanisms, at level of cortical column there should be some “sensing” mechanism able to “micro” modulate learning rate. This is occurring even in the outermost cortical levels (V1/AC1/etc…)… where i don’t think that global modulation can do the task.
Is EC/HC is an “engineering” solution to avoid dropout information? (by just copying the “suspicious” data stream that cortex regions are being ignored, and if needed, replay that stream over the cortex later).
TLDR: HTM should modulate the learning (at fine level).
Yeah nah, HTM is no where near STM. All we’re doing here is LTM. The prefrontal cortex is a whole other beast.