Hello HTM experimenters,
I have come up with a new hypothesis for sensory-motor inference. This hypothesis has two parts: one of apical dendrites and the other of the large scale structure/organization of the neural network. For both parts I will first state it and then explain it in detail. I think that this theory is biologically plausible, but I should say that my background is in computer algorithms. I am currently working to implement and test this hypothesis.
My theory is that apical dendrites are used to control HTMs. Apical dendrites have a small multiplicative effect on the proximal excitement, which max’s out at something like %2. Apical dendrites have a different learning rule than proximal and basal dendrites; instead of learning when the input predicts the post-synaptic neurons activation, the post-synaptic neurons activation predicts the input’s activation. This is the difference between P->Q and Q->P.
The pre-synaptic apical input is called the focus set (and it is of course an SDR). The goal is for the HTM to output one of the things in the focus set if one is present. Otherwise ignore the focus mechanism and output the most prominent features (which is what HTM’s normally do). This should not cause columns to activate if originally they were not near activating; this is why the effect of apical dendrites is capped at 2%. In a sentence, apical dendrites select between multiple interpretations of the data.
Example effect of apical excitement:
proximal-excitement <= proximal-excitement * min(1.02, 1 + alpha * apical-excitement)
This exact formula is guesswork, this is just a hypothesis.
The issue which this solves is that the things we are looking for might get buried in irrelevant information. Specifically, this mechanism is for the case where one area of the cortex is making basal predictions using another area’s output and wants to tell that other area to help out by looking for specific things. By this reasoning, I would expect to neurons to receive apical input from distant areas which receive basal input from the neuron. I hope this also explains why the learning rule is different.
HTMs without these apical dendrites are not controlled in any manner. They output a representation of what they see. So what happens when you show the HTM multiple things at once? For example a colored shape, which has both a shape and a color. The HTM should output a joint/combined SDR which has bits representing both shape and color, which isn’t useful. To bias it towards outputting a color (and not a shape), the focus set could be the union of all the colors.
An interesting experiment is to stare at a single point and, without moving your eyes, think of something to look for (such as the color red, or all circles, or the letter ‘E’) and observe how the things which you notice change. It helps to look an image with many objects in it. I think that this shows the extent to which your focus affects your perception.
A possible result of training an HTM using these apical dendrites is that it changes what the HTM learns. By causing certain patterns to win out in the spatial pooler, it learns more about those patterns. This could be the basis for ‘what’ and ‘where’ paths; the ‘where’ path receiving apical input from the motor cortex and the ‘what’ path receiving apical input from the control/reward areas of the brain. This explains how ‘what’ and ‘where’ paths could receive similar feed-forward proximal input but do very different things.
Figure 1, Block diagram of agent
Notice the feedback loop consisting of: processing the proprioceptive input (labeled Motor System), passing it through a spatial pooler, and using the result of the spatial pooler to directly control the muscles. In theory, this feedback loop should cause the motor HTM to associate the sensation of a movement with doing the movement. The other areas of the brain then use apical connections to control which movements it considers. User JRowe47 proposes a similar feedback loop for learning motor control.
The three main HTMs (control, sensory, motor) are connected each to all by basal and apical connections (the red arrows). This isn’t a nuanced design but illustrates that these components can and should interact with each other.
The hardwired control is where the theory runs out. As best I can tell this is where things like emotions should enter the system.
In this theory, sensory motor inference is not an explicit step. I think it will be the natural result of basal predictions and apical focus in the visual and motor areas.
I plan to test this hypothesis with an eye saccade mechanism, which searches large images for specific objects.
Thank you for reading,