Analogy of the terms 'column' and 'layer' in HTM


I read the book On Intelligence. And then, I was looking at the book BaMI. But in the Temporal Memory section, I was confused about the use of the terms column, mini-column and layers.
From On intelligence, I felt a cortical column has multiple layers (1-6). But in HTM, I felt like you were describing it as every layer has multiple columns.
Also, I would like to know whether every cell in a column (term in HTM) represent various layers as mentioned in On Intelligence.
Basically, I am confused about the exact analogy from On intelligence to HTM theory. Please explain if possible.


1 Like

We should have used the term minicolumns in BAMI. Cortical columns have 6+ layers. Some of those layers have minicolumns. See Cortical Circuitry for a graphical description and where SP/TM might fit inside a layer of a cortical column.


Thanks, also can I know the difference between a cell’s active and predicted state? I can understand it by their names. But what is the state of the cell in implementation (like does it fire - value is 1 in both the cases?)


In the biology, a predictive cell has been biased by partial activation in the dendrites of a certain layer in the recent past. This makes it ready to fire faster if a pattern that it recognizes is sensed in the proximal dendrites. This priming causes this cell to fire faster and be the winner in sensing some pattern.

So in effect - the predictive state is a bias to fire in the next processing cycle. In programming - a separate bit holds this predictive state.

Active is just the cell firing - the number of active synapses exceeded some threshold and this cell is signaling that it is active.
The pattern is recognized.

In effect, with prediction this cell is saying “I guessed this was going to happen!” For being a good predictor this cell suppresses the action of the other cells that are sensing the current input state.
Being a good guesser gets priority over being a good recognizer.

If the cell is active AND the predictive bit is true then I go back and strengthen the synapses that made the prediction.
That was a good guess.

If the predictive is set and the cell is NOT active then I weaken those synapses.
That was a bad guess.