I am requesting for any references and study material regarding voting mechanisms that the cortical columns might be using to attain a stable prediction.
I imagine this would require neuronal connections that need a considerably more proximal dendritic activation to activate the neurons in questions, if at all such connections exist. Or if modulatory connections make up the mechanism then depolarisation is probably sufficient to bias the predictions. In which case, is priming/biasing the mechanism that makes up voting?
My basic understanding is that the input space is divided into parts for individual critical columns to process and then the voting takes places in higher layers of the columns.
Any ideas or references are appreciated.
Grid formation performs this voting function.
Is that verified? Please share some resources.
Surprising how grid cells are so important. Are they present in all layers or are you referring to mechanisms that are similar to the ones in the entorhinal cortex?
There are differences in morphology in various maps (Ramón y Caja and Korbinian Brodmann noted this a long time ago) that are likely to distinguish grid-forming from sensory grouping and distributive maps. The papers that I have read put the grid-forming actions in the “hub” areas. Contrast the grid concept with the in vivo measurements in the V1 and other early visual layers.
The streams seem to function in a general plan of sharpening the sensations, then spatially spreading and intermixing the sensations to form overlapping and widely distributed representations that can be sampled for higher level grid formations.
This is “early days” in the grid field and people are just starting to look for grid-forming outside of the entorhinal cortex. Matt did post some link to papers on this in his latest HTM school video. I put up links to the same papers w/o the paywalls. I am still digesting these papers myself.
Can’t voting mechanism take place without the involvement of grid cells? Is there evidence that output from the cortical columns or the intercorticalcolumn connections end up in grid cell like pyramidal neurons or their activity pattern matches that of grid cells?
I can image cortical columns forming connections as they learn new input and these connections getting activated as new input matches the stored transitions.
I think it’s more correct to say the voting takes place across the columns. It is similar to “mini-column neighborhoods” in the SP if local inhibition is turned on (I visualized this in Topology). You don’t have to go up a layer to tally across all the layers.
I see. Thanks.
Has this been implemented in NuPIC or internal simulations?
Well yes, the spatial pooler has mini-column voting. You’re talking about “cortical column” voting, right? I think it’s going to be a similar concept.
What happens in minicolumn voting?
The way I have implemented this in my own SMI tests are that apical feed from the pooling layer gives cells in the inference layer an advantage when “predicted active” cells are selected. For example, if there are three predicted cells in a given minicolumn, and one of them is predicted both by distal and apical input, but the other two are only predicted by distal input, then the first one wins out and inhibits the other two.
@subutai In the Columns paper, wasn’t “column voting” on objects described as an effect of interlayer distal connections?
That’s fine but what about layers in the same functional level in the hierarchy. Without any top down feedback from another layer, layers should be able to converge on a stable inference.
There are long-distance connections in the pooling layer that span multiple cortical columns. These are in the same level of the hierarchy.
Do you mean connections from a pooling layer of one cortical column to pooling layer of another? These would indeed be the type of connections discussed. Which layer are you referring to exactly?
Sorry, I should have included a visualization:
In this depiction, you have three cortical columns taking input in Layer 4, and long distance distal connections in Layer 2/3a spanning multiple cortical columns. This allows the parallel cortical columns in Layer 4 to vote on what is being sensed.
(BTW, this doesn’t depict the most current theory, which has two circuits like this, one which depicts sensory inputs + orientations, and pools those into feature representations. Features are then combined with locations and pooled into object representations.)
The connections you depicted are what I mentioned here.
I am thinking about direct lateral connections between layers.
I havent read if this is how it actually happens. It might be.
But if we know the other layer exists in the cortex and theorize that it is utilized for pooling and biasing the representations in the inference layer, then do we need a solution which doesn’t utilize this layer? Even in the case where there are no long-distance distal connections in the pooling layer, the object representation in this layer would cause the inference layer to converge on stable predictions (and vice versa).
Yes, but I was looking for a voting mechanism which doesn’t involve top down feedback projection to multiple layers. I like to think that this top down mechanism is invoked by other cortical columns higher up in the hierarchy while the intercorticalcolumn voting is employed so as to converge on a stable inference before the representations travel up the hierarchy and a top down expectation is invoked.
It’s a possibility, I suppose. To me, the two-layer circuit solves the problem elegantly, so I haven’t really put a lot of thought into solving it in a single-layer circuit. I’ll leave that open for others who might have explored the topic.
Thinking about this some more, it occurred to me that a form of voting could happen even without the second layer. If we assume there are not clear boundaries between cortical columns, or if there aren’t physically distinct “cortical columns” at all (I’m not a neuroscientists, so may be completely off base here…). If cells near the edge of one cortical column were to grow distal connections into its adjacent neighbors, then a given cortical column would be influenced by the activations of its neighbors (not only by its own activations). Over multiple timesteps these influences would bleed outward, and the most consistent predictions should, in theory, win out in the end.