Whether we are navigating through city traffic or watching TV, our brain constantly generates predictions about what will happen next. Indeed, the brain has been described as a ‘prediction machine’ , building generative models that are used to predict upcoming input. However, empirical support for the generative nature of the brain largely comes from controlled but highly constrained paradigms that are a poor reflection of reality. It is therefore unclear how our predictive brain behaves when it is exposed to the vast richness and complexity of natural environments.
Recent advancements in the field of artificial intelligence (AI) have offered promising new tools to deal with this problem. Specifically, AI can be harnessed as a tool to quantify predictability in rich naturalistic environments, and provide a rich multilevel quantification of what is surprising in the input we are receiving, at different levels of complexity. In this research proposal, I will build on these recent advances in AI to parse rich and naturalistic environments, aiming to provide a computationally explicit account of the internal models that the human brain builds in the real world.
Specifically, this project studies the following questions:
- What generative models of the world do we build, when we are exposed to rich and complex naturalistic input, depending on our goals?
- What is the neurobiological implementation of these generative models? How is the bidirectional communication in the brain affected by learning generative models of the world?
- How do these generative models affect our behavior, in terms of what information we seek and what we are curious about?
Using a combination of novel generative AI tools and non-invasive brain recordings in humans, this research will bring us closer to an integrated understanding of the predictive brain in the natural world.
