@jhawkins presents ideas around how a column can learn the movements of its associated sensor and subsequently use this knowledge to move its sensor in the world.
00:00 Introduction
00:30 Subcortical Behavior Generators and Cortical Columns
16:20 Grid Cell Firing Patterns
35:40 Predictions Based on Grid Cell Theory
53:07 1D Grid Cell Minicolumns in 3D World
01:24:19 Voting and Reference Frames
Found this one really interessting. Reminds me a little bit of this talk from Barbara Webb on path intgration in ants and flies at round about 22min.
https://www.youtube.com/watch?v=m1dRO2kXNNo&t=2038s&ab_channel=TheTheoreticalNeuroscienceChannel
CORRECT WIRING
For this associative learning to work, a given CC which connects to a SCBG must also connect to sensors which sense the behavior, and to them only. The questions I’d like to discuss here about such connections are how, and why.
. Case 1: The given CC connects to only 1 sense that results from its SCBG
Here, our given CC is not interested in other senses. For example, the CC which controls a neuron in the SC is connected to only cells in the magnocellular pathway. It’s not connected to sensors in and around the eye sockets nor, for example, sensors on the kidney.
. Case 2: The given CC connects to all the senses that result from its SCBG
Here, our CC which controls a neuron in the SC is connected to both the above. Another CC, which controls a vocal chord neuron, is connected to not just auditory cells but also cells which sense vibrations near the voicebox. But, again, it’s not connected to, say, skin sensors on a toe.
The wiring in each case must be exquisite. Otherwise, the associative learning that JH suggested in this video doesn’t work.
How does such exquisite wiring happen? I can think of only 2 somewhat related possibilities:
. Evolution: The exquisite wiring is built into genes
. Pruning: The initial wiring can be much less exquisite. Connections from a toe’s skin sensors to the voicebox CC simply die from non-use
RELATED DUPLICATION
Normally, when an eye moves, both the optic flow and that eye’s socket sensors mutually agree, plus the left eye agrees with the right one. In that case, might efficiency principles prevail, therefore some sensor connections are pruned away, to save brain energy use? Or might they persist?
If they do persist, but in special cases where they don’t agree (eg. eye socket sensors give erroneous reports), our poor CC will be terribly confused.
DIFFERENT CC’s DO SAME ASSOC LEARNING
There’s another possibility: 1 CC connects to an SCBG, and a different CC connects to a resulting sense. Here, the problem is how all such CCs find one another, like friends lost in a chaotic market getting reunited, so that together they can do the associative learning linking a SCBG’s actions with the resulting feedbacks.
This would presumably happen during a baby’s first few months.
But somehow I think this won’t work.
Thanks! Do you know if there were any review paper giving an overview of the most recent theory about path integration?