A question regarding distal synapse learning

I work with a model based on biologically plausible temporal spiking neurons that use a form of STDP for synaptic learning. The model includes proximal/distal synapses, and the overall function is more-or-less the same as HTM neurons (so similar problems must be solved). I am currently dealing with the scenario described below, which is phrased in terms of the HTM neuron model (given my limited knowledge of the HTM model).

Say we are associating features with locations – as in the columns+ paper. The feature feeds a proximal synapse associated with a dendrite and the location is input to distal synapses associated with segments on that dendrite. Say a location is encoded in three dimensions, with a one-hot code for each – i.e. 3 input spikes going into the same dendritic segment sensitize the dendrite. If the given feature is simultaneously present then a segment learns the association between the feature and the location. To be specific, say the coordinates are [2, 3, 9] and say that the feature and [2,3,9] have been repeatedly input, so all the synapses associated with [2,3,9] have a permanence of 1. The neuron spikes whenever the feature + [2,3,9] are applied.

Next, the same feature is found at [20, 3, 9] – that is, two of the three coordinates are the same and the feature is the same, but the location is relatively far away from the learned location (in the X dimension). Because it is at a location far away, the new feature/location pair should ideally become associated with a different dendritic segment.

However, when the new feature/location is applied to the dendrite, two of the three distal synapses for the previous location have a permanence of 1, so the [2,3,9] segment will probably be the first to generate a dendritic spike which will lead to the neuron spiking. Without going into further detail, the new feature/location pair – which has a location quite different from the first – will not be learned. Rather, two of the three distal synapses associated with dendritic segment will be further enhanced and the two location/feature pairs in effect become associated with the same segment.

My question is: how does HTM neuron model deal with this circumstance? How can it be done in a biologically plausible way? (In a simulator there are lots of ways of doing it, but can it be done in a way that does not stretch plausibility to its limits?)


I think the fundamental issue is that some of your inputs represent a very large area of the world, and you’re relying on those same inputs to tell precisely location.

The way biology solves this problem is by using a different encoding strategy, know as grid cells: Grid cell - Wikipedia.

1 Like

I implement grid cells. The location inputs originate in the grid cells.


1 Like