Learning SDR transitions spatially via fading activations

Hello HTM theory -

I’ve been interested recently in exploring additional mechanisms that may enable sequence / transition learning, and in particular the possibility of learning transitions through the existing machinery of feed forward (proximal) spatial connectivity. There’s something tempting to me about leveraging a single mechanism for both spatial association and temporal prediction.

Based on a limited understanding of the cell level dynamics, mostly from UTHealth’s online neuroscience textbook (which seems to be a pretty incredible resource - http://neuroscience.uth.tmc.edu/ for those who haven’t seen it), it seems like fading activations could be biologically supported in certain contexts.

Without any further motivation than exploring what might be possible given a spatial pooler with continuous fading activations, I’ve been thinking along the lines below.

First add a few additional (not yet substantiated) assumptions:

  • Fade rate can be variable per cell and influenced by connectivity, e.g. perhaps cells with substantial distal connectivity in the same region fade slower.
  • An SDR X may be composed of a combination of core cells (X_c) which are highly distally connected and fade slower, and peripheral cells (X_p) which have less distal connectivity and whose activations fade faster. Some thoughts on intuition for this at end**.

Consider arbitrary SDRs A and B. A temporal transition from A -> B might look like the large grid in the lower right of the image below. Because the core cells (A_c and B_c) fade slower than the peripheral cells of A and B (A_p and B_p), A’s peripheral cells’ activations will have decreased more (or fully deactivated) by the time B is active. This leads to a resultant activation in the region with a potentially useful property: a cell in the region above could form inhibitory proximal connections with A_p and excitatory connections with B, and such a cell would detect only an explicit transition from A -> B.

To check this consider that (with the right thresholds) our cell can only fire when B is fully (currently) active, A_p has faded such that the inhibitory connections to A_p have disinhibited the cell, but A_c is still active. This can only occur in a transition from A->B.

At this point, it seems possible that with a fairly minimal hierarchy, complex and potentially high-order sequences (transitions of transitions) could be learned.

Questions on this:

  • I believe I’ve seen mention of something like this on the forum at some point in the past, but nothing came up with my recent searches. Is there a standard name for this concept?
  • Could spatial transition learning like this complement HTM sequence learning / temporal memory?

Would be curious to hear any thoughts on this, including “this isn’t worth pursuing or biologically feasible because of X, Y, Z”.


**Intuition on core vs. peripheral: One might think of core cells as composing the central, always present features of an SDR, possibly related to the concept of an invariant structure. Consider a tree. The verbal name ‘tree’ becomes highly associated with all features that are related to trees (leaves, wood, green), and so the SDR representing this word may contain many core / tightly connected cells. Whereas leaves and wood, independently, have a less direct association with each other (occur in close proximity to each other less frequently) and all other features, and so have less distal connections.


If understand it correctly, one possible issue is that this can’t learn long sequences because it is feedforward only. It might only be able to learn sequences one transition long per level of the hierarchy.

How would it represent ambiguity, where it doesn’t know the sequence context for a currently active SP cell?

As far as I know, cells are only inhibitory or excitatory, but this would require cells to be both or it would require intermediate inhibitory cells with specific targets and hebbian learning, which might not exist between levels of the hierarchy.

Maybe some sort of gradient between core/peripheral cells would help, because how closely two things correlate is on a gradient.


What you described is similar in concept to Reservoir Computing. But those tend to have fixed reservoirs with learning occurring only on the readout stage.

You might also think about your “core cells” in terms of extended activation via Metabotropic Receptors. That came up on the mailing list last year iirc.