In the third funding period, we will make extensive use both of advanced measurements of neural activity and behaviour, as well as our computational models from the first two funding periods, to study the inductive biases of visual systems that are essential for autonomous behaviour. Specifically, we will collaboratively work on the following key scientific questions and methodological challenges across all projects.
Scientifc key questions:
1. Which inductive biases and representations enable robust performance across multiple visual tasks in highly dynamic environments? In particular, (how) do visual agents use the natural structure of the world for robust performance?
2. What mechanisms underlie the efficiency of biological vision systems that enables autonomous behaviour? In particular, what is the role of active - and task-dependent - information-selection?
Methodological challenges:
1. How can we characterize the computations and mechanisms of biological vision systems with “digital twin” spanning different levels of biological plausibility? What data is needed to constrain such models, and how can they be used for mechanistic insights?
2. How can we compare intelligent visual behavior in brains and machines, in terms of representations, inductive biases and computations?
3. How can we generate a long-lasting and far-reaching scientific impact in the NeuroAI community through the provision of open-source models, evaluation-tools and data-repositories?
