TL;DR: A small shift in perspective helps interpret the ventral stream in the biological brain as the residual stream from ResNets. Experimental support, pre-, and post-dictions.
Am I really the first to comment on this interesting piece?
I'm not sure I understand the stuff about AI, but I feel I can comment on the neuroscience.
I first read about "grandmother cells" and hierarchies in the brain about 8 years ago, and have seen the arguments going back and forth since.
I have being trying to reconcile the findings from the various experiments, some of which are mentioned in this article, that show that there is some kind of hierarchy in the processing of sense data, and the requirement to create meaningful representations that the brain can use.
My conclusions are that there must be networks in the brain that represent concepts, and that any single neuron can partake in many (hundreds or thousands) of different networks. There is a balance maintained (by sleep) between excitatory and inhibitory synapse connections, which means that there is also a balance between these many networks and therefore the brain networks operate on the edge of chaos. This is why the activation of a single neuron, which could be a random event, can cause the activation of one network and the deactivation of another. This is how my attention and my thoughts flow so easily, and why it gets more difficult when I am tired.
The hierarchy of processing applies not just to sense data but also to internally-generated data from brain processes, and this is how my self-model is created, and how self-awareness comes about.
Am I really the first to comment on this interesting piece?
I'm not sure I understand the stuff about AI, but I feel I can comment on the neuroscience.
I first read about "grandmother cells" and hierarchies in the brain about 8 years ago, and have seen the arguments going back and forth since.
I have being trying to reconcile the findings from the various experiments, some of which are mentioned in this article, that show that there is some kind of hierarchy in the processing of sense data, and the requirement to create meaningful representations that the brain can use.
(In passing, Jeff Hawkins' https://en.wikipedia.org/wiki/Memory-prediction_framework may be of interest.)
My conclusions are that there must be networks in the brain that represent concepts, and that any single neuron can partake in many (hundreds or thousands) of different networks. There is a balance maintained (by sleep) between excitatory and inhibitory synapse connections, which means that there is also a balance between these many networks and therefore the brain networks operate on the edge of chaos. This is why the activation of a single neuron, which could be a random event, can cause the activation of one network and the deactivation of another. This is how my attention and my thoughts flow so easily, and why it gets more difficult when I am tired.
The hierarchy of processing applies not just to sense data but also to internally-generated data from brain processes, and this is how my self-model is created, and how self-awareness comes about.
I have documented full details of my proposals on a new website: http://www.hierarchicalbrain.com