In recent years, neuroscientists have gotten keenly interested in behavior. Advanced video tracking methods and AI have allowed us to follow the detailed movements of animals. Then we can use sophisticated behavioral modeling to divide that behavior - in an unsupervised way - into discrete states, and do exciting neuroscience with it.

​

This is the ethogramming approach. The name comes from field biology, which uses similar approaches, although done laboriously, by hand, by humans. That is more or less how we have formalized behavior for 100+ years. Divide behavior into discrete categories - groom, fight, forage, and string them together, serially, like beads on a bracelet.

​

So, from a scientific perspective, the order of operations is observe, divide, label. It’s the bottom up approach to understanding animal behavior. This is a really useful approach and it’s told us a lot. Our lab has done it too.

​

But there’s something missing from the whole approach. It misses how behavior is generated, and as a result, the observe-divide-label approach misses something about how behavior works. 

 

Because we don’t really generate our behavior like beads on a string. Really, what we do, is we have goals. We have things we want. Those come first. The goals may change over time, but it’s the goals at any moment that drive what we do. And the behavior is second; it is downstream of our goals. 

 

So the goals are at the top. And a top-down approach to understanding the neural basis of behavior would be one that seeks first to identify the goals, then infer how the goals drive the behavior.  

 

This makes a huge difference. That’s because goals are not all-or-none. We might have two goals at the same time. We might have a hierarchy of goals. One goal may be waxing and the other goal may be waning. A bottom-up approach is blind to all of that because it forces behavior at any time into one category - it’s either groom or fight, or forage. Never a blend. 

​

So in this paper, we argue that the next step in neuroethology is goal inference. And that's harder than current unsupervised approaches, which are basically fancy clustering in the high dimensional space of behavior. As scientists, we are blind to goals, and they can be tricky to infer from behavior. Is that dog running after a prey or running away from a predator? Even if our subjects are humans, they often don’t know their own motivations, and if they do, can’t alway express them. So to understand behavior, we as scientists have to do latent goal inference.  

​

In neuroeconomics, we are used to thinking of choice as an all-or-none process. But in the real world it seldom is. Typically, in the real world, we can try to have our cake and eat it too, even if carefully designed laboratory experiments try to preclude that possibility. But we think neuroscience needs to open that door again. We need to study choice as a continuous process, one that can involved blended strategies, not just discrete choices.

 

Just as microeconomic theory is a good foundation for discrete choices, control theory is a good foundation for continuous choices.

​

Our paper is about a really simple task experiment that doesnt have any grooming, any fighting, or any foraging, except in the simplest sense. It’s as absolutely stripped down at it can possibly be. As simple as possible, but no simpler. 

 

It’s the prey-pursuit task, basically, an ultra-simple version of chasing, the same game kids and dogs love so much. The player uses a joystick to move an avatar on a rectangle pen, and goes after a moving prey. The prey has a tiny amount of AI (for video game nerds, it’s called A*) to evade the pursuer. On some trials there’s a predator. 

 

We work with epilepsy patients at Baylor St Lukes hospital, people with a large number of electrodes implanted in their brains as part of a treatment for serious epilepsy. The surgeons place the electrodes, and it depends on where the neurologists think the epilepsy is coming from, but very typically, it involves electrodes in the hippocampus, anterior cingulate cortex, and orbitofrontal cortex. 

 

As part of their treatment, the patients have a hospital stay where their brain activity is recorded 24 hours a day, 7 days a week. We will sometimes ask for half an hour of their time if they want to play a simple video game kind of like a simpler version of Pac-man.