There is a gap in my understanding of how the activation of cells happens biologically. I hope somebody is able to provide clarifying insight.
Let us say there exist learned sequences ‘ABC’ and ‘FBD’. When A columns fire, specific cells in B columns become active. When, after that, B columns fire, specific cells in C columns become active, but not in D columns because B was fired in the context of ‘AB’. My question is, under which exact circumstances and how does a cell get activated?
The reason I ask for the clarification is this seeming contradiction: if the rule is “the firing of a column activates cells strongly related to that column”, then after firing columns of A and B, predictied cells in columns C as well as D would activate, because they are both part of learned sequences following B and thus strongly related to B.
However, if the rule was “the firing of a column which had a predicted cell, activates those cells related to its correctly predicted cell”, then columns C would activate.
In NuPIC, both of these rules are used (right?). The latter is used unless there was no correctly predicted cell. In that case, the former is used.
But how could the biological system possibly make this distinction? Biologically, you would expect that, if the firing of a column can activate all strongly related cells, it would always do so. Concretely in this case, if B is seen without context, it activates strongly related cells of columns C and D.
So, a correctly predicted cell would have to shut off the rest of the column or something? As far as I understand, it doesn’t add up entirely.
Thank you for your reply, and the excellent video on temporal memory.
I apologize for the mix-up of terms. (I used ‘active’ where it should have been predictive, and ‘fired’ where it should have been active)
Oh, not the entire column becomes active, only the cells that were in a predictive state.
But the input always makes the entire column become active, because it is connected to the column and not individual cells in the column, right? That seems contradictory.
Does the entire column being activated by the input not imply the same consequences as bursting? An active column is an active column, so active cells in the column should act the same way unless another mechanism acts on them. I can see the cell in a predictive state acting differently, but not why the cells in an unpredicted state would act differently than if the column was bursting.
To make it logically sound, either a mechanism should prevent cells other than the predicted one becoming active, or something should prevent other active cells in the column to instigate a predictive state.
There is a difference between an “active column” and an “active cell”. Spatial Pooling provides a stream of active columns representing spatial information. The TM algorithm operates within that space. It only activates cells that are within active columns. You’ll never see a cell activate in HTM outside of an active column. This means that the proximal input is the ultimate driver of cell activations.
I don’t know the exact neuroscience of how proximal input is shared by the cells within a column.
I explain the two conditions that cells become active at 5:32 in the video.
When a column is activated, as determined by the spatial pooler, one of two things can happen with regard to the cells inside the column. If there are no cells in the column that are in a slightly depolarized “predicted” state, all the cells within the column become active together (bursting). There have been observations in biological experiments that cell firing activity becomes more dense in the presence of unpredictable patterns versus repetitive, predictable patterns. Alternatively, if there are any combination of cells in the column that are in a predicted state, that subset of cells become active first and inhibit all of their neighbors in the column.
