I was watching neuroscience video timestamped link, where it says:
Neurons preferentially connect to and form strong synapses with other neurons responding to lines with the same orientation, whereas inhibitory interneurons innervate surrounding neurons regardless of orientation tuning.
The video is saying that neurons which receive the same receptive field input tend to connect to each other (excitatory recurrent connection) to overcome a blanket of inhibition.
The HTM theory suggests the opposite, that neurons which receive the same bottom-up input want to inhibit each other so that only a neuron which has the current context state fires.
Is one of these statements incorrect, or are they both somehow compatible. For example, is it the case that neurons which are receptive to the note E, and predict themselves based on the note G, are likely to wire together to overcome a blanket of inhibition?
The neurons that are reinforcing each other have extra excitement vs the ones that don’t.
Everyone that is receiving “something” triggers the inhibitory field but the cooperative cells are better able to assert themselves in opposing the inhibitory field.
In practice, this works somewhat differently that the stock K-means that is part of the stock HTM model.
Connections between neurons with the same orientation preference could be 2 things.
It could be cells in the same minicolumn. That makes sense for generating a shared orientation preference (which HTM models as a single proximal dendrite), because it causes the cells in a minicolumn to tend to fire at the same time and so learn the same feedforward input on their proximal dendrites.
It could also be cells in different minicolumns with the same orientation preference. That means different cortical columns (macrocolumns), which are outside the scope of the spatial pooler and temporal memory. Those things don’t have multiple cortical columns, and also don’t have strong lateral excitatory synapses. (The temporal memory has lateral synapses, but to produce a measurable signal at the cell body, multiple synapses on a dendritic segment must activate.)
TBT is much more theoretical, so exact mechanisms aren’t nailed down. I think the output layer could have lateral synapses which’d individually produce a measurable signal at the cell body. If the object is a line at a certain orientation, the output layer’s neurons for that object would be connected (for voting).
I think these are two different types of lateral connections: short-range inhibition and long-range excitation, AKA Mexican hat model: The Influence of Mexican Hat Recurrent Connectivity on Noise Correlations and Stimulus Encoding - PMC
I don’t think HTM is currently modeling the second kind, please let me know if that’s wrong.
Thanks, I didn’t know this was the case.
I guess it’s possible that similarly activating neurons could make longer-range synapses with Hebbian learning (the 2nd kind), and the short range inhibition could do the context association of sensory input.
That video is full of oversimplifications. The things are much, much messy and harder. Don’t get fooled.