Quick question, I’ve watched all the HTM school videos. I noticed in all the ideas of the SDRs, especially the union videos, the focus is on 2-dimensional SDRs, why not 3d or n-dimensional SDRs? Was there a specific reason to limit it to 2d?
I chose 2D for visualization reasons only. The same concepts apply to any dimensions.
Maybe this is a stupid question, but other than make an easier visual representation, why would an SDR be anything else than unidimentional? As far as I understand, none of the bits’ significance are related to their location in the SDR. Or am I missing something important?
When we talk about dimensionality, we’re also talking about topology. Sensory input is not unidirectional, so as it is converted into SDRs, you can’t expect the SDRs to be unidirectional. It makes it simpler for us, but I’d venture to say that all sensory input to the brain is topological, thus has more than uni-dimensionality.
I’ve been thinking about this. Has it really?
I mean, I understand that in some (if not most) area’s in our brain it is like that, because the signals happen to enter in specific topographical shapes. It’s logical that our neuronal hierarchy grew in that sense.
But is it a prequisite for HTM? Is it even an advantage?
I have a few arguments:
If we want HTM to work uniformely no matter where a particular column is located, the column needs to be able to interpret information without caring for the topology of the input. And so must its components.
The information from our retina runs over bundles of nerves that snake across our brain. Do the endings connect with the same topology into the visual cortex?
If topology was important, wouldn’t recognising the same object in different scales be a difficult task? (Moving towards an object would be a weird experience :-). ).
Abstract concepts would be very hard to represent.
I wonder how much our cognitive abilities would improve if neurons were not limited to the number and origin of other neurons due to spacial constraints.