I think it wasn’t mentioned yet [2412.18354v1] The Thousand Brains Project: A New Paradigm for Sensorimotor Intelligence
2 Likes
I think a potential hole in this model/architecture (and many others) is that it skips, (doesn’t observe, blindisghts) a fundamental step.
Here-s what I mean:
One core component in TB project is the learning module with the purpose to learn and later recognize an object (or more likely a class of similar or related objects)
The paper follows up with an example object/experiment in which the model is subjected to interactive observations & manipulations of the proverbial cup, and hopefully succeeds in learning what a cup is and later recognize it in various contexts in which a cup happens to be visible or touched.
That approach, while conceptually makes sense it skips an important, lower level problem that needs to be solved before even considering to learn an object.
That problem is simply: How on earth does the tousand brains even consider or manage to figure out that there-s something in their receptive/sensory stream that is supposed to be considered as an individual object that needs to be segmented out off that sensory stream(s) and investigated in order to learn anything about that object? And its properties, behaviour, poses, etc. that will later hint “woops, there-s a famous cup”
TLDR how does an object pops out into awareness of the brain?
1 Like
I know a theory about how to do this.
- Here is a paper explaining it: https://www.researchgate.net/publication/215991433_Learning_Invariance_From_Transformation_Sequences
- Here is my lecture explaining it: Video Lecture of Kropff & Treves, 2008
I hope this helps
1 Like