Understanding human intelligence and building strong AI systems is a key challenge for our generation. A particularly puzzling aspect is that the human brain seems to cope very well with the highly variable and uncertain nature of perception and action, regarding both their signal characteristic and how they extend over time. Furthermore, it seems apparent that the information processing in the brain always involves previous bodily experiences and all our senses, thus is embodied and crossmodal. For explaining these characteristics, there are strong accounts that the brain is constantly predicting sensory input and feedback while minimising free energy in the prediction and is hierarchically abstracting the perception as well as hierarchically composing action. Furthermore, it is also hypothesised that mechanistic priors in the brain’s information processing induce a failure of hierarchical inference in the brain, accounting for atypical perception and action of psychiatric disorders. For example, some brains might have developed to focus too strongly on current sensory input while others might focus too strongly on memorised previous experience. A typically developed brain, in contrast, would show a fine balance compared to these two extreme priors. The big open mystery is: how is the brain developing this on a mechanistic level and thus how can this get learned within an AI system?
Thus, to lift this mystery, we further need to bring together research from computational neuroscience, cognitive psychology, and artificial intelligence. With this workshop, we particularly want to wrap up different recent hypotheses, models, and experiments and discuss in-depth how to shape future imaging, behavioural, and developmental robotics studies as a complement to computational modelling and bio-inspired artificial intelligent algorithms. As a guiding theme, we aim to approach the following central questions:
How does the brain learn spatio-temporal stochasticity in perception and action?
What is the role of priors in learning spatio-temporal adaptive prediction?
How can developmental robotics help us to study spatio-temporal stochasticity as an analogy to infant learning?
We examine these questions with the insight from invited key speakers from complementary fields as well as original contributions in breakout sessions in order to conclude the next steps within a panel discussion.
For this we invite submissions of extended abstracts (one page including references) on ongoing research that focuses on one or several of these questions as contributions. In addition to the live presentation and discussion during the workshop, the contributions will be made public in workshop proceedings.