Biased Predictions inherent to HTM?



I was thinking about how the hierarchy works…

When a higher region sends down a prediction of what the lower regions will see, it is biased to align with predictions it received from the region above it. right? In other words what it sends down to layer 1 in the region below is a representation of two things: what the higher-ups want to see happen (the goal) combined with what the higher-ups expect to see happen (the prediction).

Thus data flowing down the hierarchy can be thought of as predictions, biased by goals.

But what about data flowing up the hierarchy? What are those predictions biased by?

When a lower region sends data up to the next level in the hierarchy is it’s data biased by a goal or is it biased by what it knows it’s lower regions, and itself are planning on doing? Is it biased by its actions, it’s behaviors it’s already sent out to the body?

Am I understanding the HTM algorithm correctly?

Flowing from the bottom up is there any bias towards any goal that gets added to the prediction? I know a lower region will try to interpret its data in alignment with what it received in its layer 1 but is that the only goal-based bias inherent in its sent up prediction?


Hi Jordan,
The quick answer is, no one knows how to interpret hierarchical processing in the cortex. We have many clues, but not a definitive theory.

Until recently, I had espoused somewhat conventional views of hierarchical processing, with the addition of temporal consolidation as information ascends the hierarchy. Hence, I came up with the term Hierarchical Temporal Memory. However, in the past two years we have made much progress in understanding what cortical columns do. We now realize that a single column (or a set of them in a cortical region) is much more powerful than anyone realized. A single column and region can learn complete models of complex objects. This changes how we think of hierarchy.

In a recent talk at MIT I discussed our latest understanding of what cortical columns are doing. I had a few slides toward the end titled “Rethinking Hierarchy” where I suggested some of the implications for hierarchy given our new understanding of columns. You might want to look at it to see where my current thinking is. Here is the link

The bottom line is, at Numenta we continue to think about hierarchy all the time. But, we don’t think we can understand it in detail until we know exactly what each region/column is doing. We should be there shortly.