Study reveals hidden brain circuit behind flexible visual thinking

To test whether that wiring was essential, they weakened those connections in the model, and its ability to switch between tasks collapsed. Weakening other types of connections left performance largely intact. The pattern held up against the living brain as well. In recordings from the visual cortex of mice, silencing the inhibitory cells that anchor this circuit reduced the cortex's ability to track the task context, just as the model predicted.

This builds on last year's fMRI study. Why turn to AI models next?

A brain scan gives you a picture of the whole brain, but it's coarse - you can't see what individual cells are doing, or what's happening at the level of circuits. Seeing which regions light up and patterns of activity gave us a reason to go deeper. To get at the mechanism, we needed something we could take apart and change, so we turned to neural models. We build these from scratch, so it's possible to see exactly what the network is doing to solve the task.

Why keep the models so simple?

If a model has abilities the brain doesn't have, then anything we find inside it won't tell us much about real brains. So we did the opposite and built something that only includes features we know to be true about biology. A lot of that builds on earlier work from Tomas Gallo Aquino and Robert Kim, the paper's co-first authors, among other studies. We know there are excitatory and inhibitory neurons, so we built those in. We know the brain is organized in a hierarchy, so in the second part of the paper we gave the network two regions: a sensory module that receives input directly, and a higher-level module downstream. Then we can ask how the network uses those ingredients to do the task.

Why do these inhibitory-on-inhibitory connections matter so much?

They give the system very fine control over how information gets represented. These inhibitory neurons turn out to be really important for keeping everything well-controlled, for making sure the right thing is represented in the right way. There are four kinds of connections you can have between these cells, and the one that matters for this kind of flexible processing is inhibition acting on inhibition. We know it's important, because the model fails when we take it away. We don't yet know why it has to be this particular wiring. That question is an important topic for research teams across the world.

How did you get started with this line of research?

Back in 2015, I began working with patients who are missing the hippocampus, the region that lets us form and hold onto memories. If the brain were truly modular, losing that middle piece should leave you unable to do a great many things. But it doesn't. These patients can still do all sorts of tasks. That was the first real evidence, to me, that the early regions of the brain are doing more than just relaying sensory information. And it points to something useful: if information is stored redundantly, you can lose a part and still get by. I think that's how the brain actually works.

What could it mean for AI?

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.

What's next?

We've gone back to humans. We're working closely with clinical collaborators who monitor epilepsy patients with electrodes placed deep inside the brain, letting us record neural activity directly while those patients perform cognitive tasks. These fine-grained measurements will give us data to test our hypotheses against real neural activity.

This work was funded by the ARL Human Guided Intelligent Systems grant (W911NF-23-2-0067) and the Strengthening Teamwork for Robust Operations in Novel Groups (STRONG) grant (W911NF-22-2-0148).

Source:

Columbia University School of Engineering and Applied Science