The circuit that lets your brain think and see

Paper title: Disinhibitory signaling enables flexible coding of top-down information in cortical networks

(should be qualified as in-silico visual systems)

Method: replicate fMRI findings of visual abstraction using simple networks to model what's essential

Gist: in tasks 'Inhibitory neurons that suppress other inhibitory neurons seem to pass key information from the “thinking” part of the system to the “sensing” component of the system'

I've heard the same for motor control: it's not that the cortex aims for one action; it aims for a bunch, but most are inhibited. (You see this in chaotic movement when inhibition fails).

So it's not really "think and see" but "what you see when you're doing a task".

(There's some analogy in there wrt (AI) exuberance effacing selectivity in investment decisions...)

The actual experiment is basically training a relatively large RNN (1000 units) to do a very simple copy-esque task. The weight values are constrained to be either positive or negative at the beginning of training. The RNN could probably be way smaller and still be made to solve this task.

It isn't even a spiking model.

It seems really hard to go from this experiment to "we've learned anything useful about how brains actually work."

https://www.biorxiv.org/content/10.1101/2023.10.17.562828v2....

Why are they using neural nets to model observed behavior (different parts activated) and then applying them to biological neurons that work completely differently? Real neurons communicate using precisely timed spikes and each neuron does a bunch of local computation as well.

Independent of the research itself, the article makes it seem as if neuroscientist are just discovering the deep recursion all the way back to V1. The idea that this was a one way stream of information processing was discarded a long time ago. Those back projections probably serve lots of functions, but we can be pretty sure they are there to let current context bias the weights for quicker recognition and reaction...e.g. if your context includes snakes, your visual system will attune to recognizing snakes even faster.

That’s…quite the epic surname.

"Compare the brain to something like ChatGPT or a large language model. We can do far more, across far more situations, on a tiny fraction of the energy — and without being trained on the whole internet. The brain got there through evolution, through the redundancy built into its wiring. Our models are recurrent neural networks, which are quite different from the transformers behind today's large language models. The goal is to work out these principles one by one and use them to make AI leaner and more adaptive. This inhibition-on-inhibition motif is one of them."

Cool!

[dead]

[dead]

Reverse engineering how algorithms in the brain work is a really promising path towards making genuine AI systems which would make the current crop of LLMs obsolete.