This is a recent strand of thought and experiments in the lab. Most neuroscientists alive want to understand the relation of neural circuits to animal behavior. While there has been an explosion of new methods for observing and manipulating neural circuits, nothing comparable has happened for the other side of the equation, namely animal behavior. In the case of the house mouse - a popular species for studying neural circuits - we know rather little about its natural behavior. There exists exactly one book on the topic: "Mice all over" by Peter Crowcroft (it is excellent). So it is difficult to mount a "neural-circuits-to-behavior" program as long as we don't know what behaviors need to be explained.
Many contemporary studies solve the dilemma by making the poor mouse perform some abstract task and calling that "the behavior". A popular choice is to bolt the animal into a head-fixed setup (convenient for recording or imaging) and making it learn some two-alternative-forced-choice (2AFC) rule. For example, a light appears on the left or the right and the animal has to bend its tongue towards the side of the light. As a rule these "behaviors" are unimpressive. The animals take many thousands of trials to learn this, and they never get particularly good at it, with a typical error rate of 20% on a truly trivial question [1]. Not the sort of performance you would actually need a brain for. At the same time, we know that mice can learn to play basketball in a few days. So there is a major disconnect between the so-called "behaviors" in much of the mouse neural circuits literature and what mice are actually capable of [2].
On that background we want to develop ways to let laboratory mice exhibit more complex skills. Of course in a laboratory setting that allows accurate measurements of their behavior and the neural signals that contribute to it. Our first stab at this involves a tool used by animal psychologists over a century ago: the complex labyrinth [3]. When paired with modern methods of animal tracking and computational modeling this may offer a new view of learning and decision-making in biological agents.
One of the insights from the labyrinth experiments is that mice can learn a complex sequence of actions (e.g. 6 correct 3-way turns) after just a few reward experiences. We are exploring how such rapid learning on a decision graph could be implemented with biologically realistic circuits. One such proposal is an algorithm we term "endotaxis" [4]: The underlying model circuit allows the agent to learn both a map of the environment and the location of interesting targets. Subsequently the agent navigates by following a "virtual odor" computed by the neural circuit that leads it to the goal.
[1] Qiao, M., Zhang, T., Segalin, C., Sam, S., Perona, P., and Meister, M. (2018). Mouse Academy: high-throughput automated training and trial-by-trial behavioral analysis during learning. BioRxiv 467878.
[2] Meister, M. (2022). Learning, fast and slow. Current Opinion in Neurobiology 75, 102555. 10.1016/j.conb.2022.102555.
[3] Rosenberg, M., Zhang, T., Perona, P., and Meister, M. (2021). Mice in a labyrinth show rapid learning, sudden insight, and efficient exploration. eLife 2021;10:e66175. DOI: 10.7554/eLife.66175. For easier reading check out our preprint on BioRxiv.
[4] Zhang, T., Rosenberg, M., Perona, P., and Meister, M. (2022). Endotaxis: A neuromorphic algorithm for mapping, goal-learning, navigation, and patrolling. bioRxiv https://doi.org/10.1101/2021.09.24.461751