Sorry I have to do this one at 9AM Pacific because I’ll be OOTO in the afternoon.
I will be talking about:
What do you want to discuss?
Is there any chance that we can do deep and discuss how location modules work(from a biological point) and how does the SDR it generated fulfill the description you gave on GitHub? I have been reading the paper and porting location_module.py to C++. But I still can’t grasp the internal mechanism.
Yes I will talk in detail about this, and we’ll continue conversation in the 2D Object Recognition Project thread.
Is there a path to work with TBT concepts using canonical HTM tools?
Yes, I followed the references to the paper:
I was looking for a bridge to add in the TBT ideas.
If you are talking about the lateral connections between the object layers, doesn’t the supporting code for the Columns Paper give you that bridge? This is at least the reference I was going to use when we got to that point in development, but there is more groundwork to do first.
I am trying to see where I differ from the Numenta tools. This might help me understand why we seem to be on different pages when it comes to how the lateral connections are working. Having actual code to look at would go a long way to explaining the vision of how it works.
I will give the link you provided some attention and see if I get my answers; failing that - I’ll wait for you to do the groundwork and presentation.
I also want to talk about python 3 and NuPIC 2 (added it to the list). This hangout is happening live in about 40 minutes. Will post the link to join soon.
Join here anytime. I’m early, just setting things up .
One of the things that jump out to me going through it is that the Numenta simulation does lateral connections to ALL cell bodies.
You may recall that I consider the length of the lateral connection to be an important consideration as it clearly is in biology. Topology matters. This leads to the hex-grid signaling in my models.
Please consider the implications of hex-grid coding and how this can build to the observed properties of Moser Grid cell. What do I mean by building to Moser grids? Consider this image.
The red and blue hex-grids are formed in separate maps and project to a third, say the EC.
These hex-grids are formed by hex-grid coding whatever stream of information that they are processing. I have shown the entire extent of the hex-grid but it is likely that this would be a much smaller patch of hex-grid activation as the stimulus to the map changes.
This is the sparse and active cells in the L2/3 layer and the output axons project to other maps.
In the map where they join there are hot spots formed as the information that they are coding intersects and reinforces. (the yellow patches)
Ignoring this biological property of lateral connections is why this is a big yawn to Numenta.
It is critical to my understanding the inter-map interactions.
If you look for this type of pattern, you’ll find it in many places in nature. I’m still not sure it is important for computation in the brain.
For the pattern to emerge there has to be some underlying periodic structure.
I don’t see how this can happen with random interconnections.
I have posted elsewhere that there is a preferred lateral interconnections length. This is sufficient to establish the structure and has the additional charm that it is a local feature that does not require any global mechanism after the initial grow-into-place structure during neurogenesis.
Quick clarification on my point about vision probably not being the best example to use when talking about the object-centric orientation question. In the visual system attention is an extremely important factor that must be taken into account when when doing thought experiments related to images (I may not notice an eye being flipped simply because I am attending to the face, not the eye). Additionally, the amygdala, for example, is programmed with important shapes like faces, and it is also an input to the cortex, and is built into the “face” models the cortex is constructing.