Today in like 30 minutes.
4 Likes
@vlomonaco is talking about Animal AI Olympics.
1 Like
@mrcslws is not talking about two types of SDRs (semantic vs symbolic).
The slides about the Animal-AI Olympics part are available here!
2 Likes
This was a great stream again.
At 01:15:22 @jhawkins talked about something I haven’t heard before, but whiich sounds very important:
Does anyone know more about this? Is it part of a Numenta paper I missed? Can someone explain, please.
2 Likes
This goes to what I have been saying about local hotspots of activity.
I am still looking for confirmation or falsification that more than one spot of activity can exist at the the same time in the same map. If so - I think that this could go a long way to the number of simultaneous things that you can have going on at the same time.
1 Like
I think the bumpiness might have something to do with the continuous distribution of selectivity among the columns (e.g. edge-orientation selectivity in V1). IIRC the population encoding scheme utilizes correlations between multiple Gaussian bumps over multiple sensors and/or grid cell modules to produce a representation that is actually able to recover a much sharper resolution on the encoded signal than any of the individual noisy encoders would be capable of on their own.
The bleed over from one column to the next might just be a side affect of those noisy sensor elements attempting encode some kind of probability distribution over the population. It then takes multiple simultaneous observations with similar elements combined with cross-column voting to generate a consensus on the actual encoded value (i.e. wisdom of the crowds).
Such an encoding scheme is totally consistent with the hypothesis that grid-cell-like modules are responsible for not only encoding an internal representation of parametric variations of an observed property, but also allowing an agent to mentally navigate through that very same parametric space.
2 Likes