A segment that gets predictive at “Hello” and activates after “how are you” is encountered might be useful if it isn’t prematurely punished.
Leave a decrementing time step counter to delay punishment.
In the fortunate case it does activate in few time steps, then THAT could be an useful new pattern.
Whoever happens to fall within the hole (George, Nancy, etc) might wait on another delayed segment a couple time steps to be offered its surrounding context. And the holey segment makes an interesting generalization since it doesn’t care who is greeted.
Some segments persistently activating a certain window forward could be given a permanence score for the predicted time steps.
So they can be marked as “delayed” till the time gap passes, to not clutter the list of currently expected activations.
I am (very slowly) learning Spanish.
One of the things jumped out to me is that word order is very different and the “hole” may be very far away from where it might appear in English. It seems very odd to me as a native English speaker.
What I am getting to is just this: I think that in human language production & recognition there are two primary areas - one in the motor planning area and one one in the border between the temporal lobe and parietal lobe. One of these areas is heavy involved with word sound/shape and meaning. I think of this as being similar to the cortical IO retina. The one in the frontal lobe is concerned with grammatical structures and learns templates for allowed grammar structures. This should not be a surprise as it could be considered as part of motor planning. The two areas are connected with hefty fiber bundles and I assume the work cooperatively for both recognition and production of speech.
What does this have to do with training basal dendrites?
Some parts of pattern learning are associated at “higher levels” of pattern distribution. For example; one of the more successful things drawn from modeling aspects of the brain is deep learning. The usual training method (back prop) is not biologically inspired but the application of distributed data (both longitudinally and laterally in a hierarchy) allows what I call “micro-parsing.” That is the decomposition and aggregation of meaning between levels in a hierarchy. The success of deep learning offers the possibility of huge improvements in HTM function when the H is finally addressed.
And how does this apply to the Basal Dendrites thing?
Perhaps it is important to reflect on the best place in a system to place a function; it may not be at the local dendrite level but instead to the larger structure it is part of.
This changing of word order is an interesting characteristic when you look at the full range of permulations used around the world… they may share one interesting characteristic though,
Japanese (SOV) “She loves him”
Arabic (VSO) “Loves she him”
Fijian (VOS) “Loves him she”
Hixkaryana (OVS) “Him loves she”
Xavante (OSV) “Him she loves” (and Yoda)
When you look at the characteristics of the ordering and how they group on a conceptual basis it’s quite interesting.
Why not attach a time reference within the HTM model as a whole (every synapse equivalent) and provide proportional punishment ?
When we would learn the sequence you indicate “Hello [xxxx] how are you ?” we would first learn “how are you ?” separate to “Hello” So the pattern would really be “Hello [xxxxx] [prior learnt pattern]”
Buffer the input and allow for a recursive integration so that “how are you” can be identified as a separate pattern and then fed back in… ?
The holes of which you speak, could also be interpreted and modeled as symbolic place holders.
We have the capacity to recognize idiomatic sequences in languages as chunks of content that have an associated meaning. We also have the ability to recognize portions of sequences that can take on nearly arbitrary values without significantly altering the encoding if the rest of the sequence. Sometimes this placeholder can be blanked/bleeped out or omitted entirely, and the meaning of the rest of the sequence remains intact. Often the missing value can even be implied from other contextual cues.
My current conceptual model places each of these tokens in a series of nested loops along with the rest of the contents of consciousness. Contents with greater saliency and confidence serve as landmarks to anchor the current context. As the loop plays out over several cycles, other tokens, with less saliency or confidence, are gradually filled in or replaced with concepts that are highly plausible given the current context. Each of the nested loops provide a portion of the context used to resolve any remaining ambiguities.
I suspect that these same loops could also be used to evaluate the potential saliency of hypothetical scenarios - a sort of what-if game playing with variations on a theme.
While all linguistic considerations mentioned are interesting, all I wanted to say is a relatively simple change in TM algorithm would make it more plausible biologically and have the useful consequence of once sequence AB is learned, it will respond, within a limited time window to patterns like A_B and A___B too.
Sorry for the linguistic example I started with.
@BrainVx Partial penalties I’m not sure is right. Delaying a configurable number of time steps the decision on whether to reward or penalize seems more “correct”, biologically plausible and simpler.
@CollinsEM - those are very interesting ideas but they-re targeting a different complexity and time scale than the 10-50ms in which a single TM cell is predictive. They would make an interesting separate thread.
PS
Hopefully a more convincing example:
Let’s say a set of cells/columns learn pattern ABC, another learns XYZ.
Each of the two patterns will be recognized correctly when the two sequences overlap, e.g.
XABYCZ
By not allowing “gaps” between the time of prediction and that of activation there are lots of possible combinations to be learned and each would hardly point towards the two initial simple patterns.