Final Report Summary - RATLAND (Understanding Auditory Information Processing in Naturalistic Environments)
We developed a unique behavioral setup that allows us to study animal behavior and associated brain activity in a complex environment under tight experimental control. To achieve these goals, we developed on the one hand a setup (the RIFF - rat interactive fantasy facility) that allow us to follow animal behavior at a high precision and collect synchronously brain activity, and on the other hand we developed a set of theoretical concepts and computational tools that make it possible to interpret animal behavior within the RIFF and to link it with brain responses.
The RIFF implements a model of the interaction between an animal and its environment that is called a Markov decision process (MDP). MDPs are simple, yet powerful and flexible models of the interaction between agents and their environments, and are used extensively in fields such as reinforcement learning. Importantly to us, MDPs often have optimal policies - a way for an agent to maximize reward through its actions. However, rats do not follow optimal policies. To account for this, we used a novel concept, that of optimal policy under information constraints. In short, optimal policies may be rather complex, requiring the animal to have a highly precise idea of the state of the environment (requiring high perceptual information flow) as well as making decisions that are highly precise (requiring high information flow of actions into the environment). We developed ways of finding policies that are optimal under constraints on the complexity of the policy, and found that animal behavior is described much better by such policies. This work provides us with a microscopic view of animal behavior - a detailed description of the rules by which it selects action on a moment-by-moment basis. At the same time, we record brain activity, that follow the moment-by-moment internal representations and decisions of the animal. Our computational tools allow us to produce a joint representation of animal behavior and brain activity, the infrastructure required for understanding the links between the two at an unprecedented level of detail.
The flexibility of the RIFF allows us to pursue a number of different specific sets of questions. We use the RIFF to study auditory perception - for example sound localization by rats - as well as decision making and habit formation. At the same time, we can record brain activity from multiple brain areas, from purely auditory areas to high-level integratory centers such as the insular cortex.
The RIFF implements a model of the interaction between an animal and its environment that is called a Markov decision process (MDP). MDPs are simple, yet powerful and flexible models of the interaction between agents and their environments, and are used extensively in fields such as reinforcement learning. Importantly to us, MDPs often have optimal policies - a way for an agent to maximize reward through its actions. However, rats do not follow optimal policies. To account for this, we used a novel concept, that of optimal policy under information constraints. In short, optimal policies may be rather complex, requiring the animal to have a highly precise idea of the state of the environment (requiring high perceptual information flow) as well as making decisions that are highly precise (requiring high information flow of actions into the environment). We developed ways of finding policies that are optimal under constraints on the complexity of the policy, and found that animal behavior is described much better by such policies. This work provides us with a microscopic view of animal behavior - a detailed description of the rules by which it selects action on a moment-by-moment basis. At the same time, we record brain activity, that follow the moment-by-moment internal representations and decisions of the animal. Our computational tools allow us to produce a joint representation of animal behavior and brain activity, the infrastructure required for understanding the links between the two at an unprecedented level of detail.
The flexibility of the RIFF allows us to pursue a number of different specific sets of questions. We use the RIFF to study auditory perception - for example sound localization by rats - as well as decision making and habit formation. At the same time, we can record brain activity from multiple brain areas, from purely auditory areas to high-level integratory centers such as the insular cortex.