Artificial Intelligence has impacted across science and society, but the basic mechanisms and structures are still poorly understood. While many machine learning models require large datasets for learning, the brain generates its own data through exploration, mental replay, observing others and from patching partial observations. We are able to create possible future scenarios that are mechanistily plausible, that adhere to the cause-effect principles that are found in nature. However, this relies on having the ability to intervene with structure and laws in the world.
Fundamental aspects of space and relational reasoning must be learned from experience as illustrated in Fig. 1 A. Here, an agent (represented by a rat) is presented with sensory observations such as a spider web or a squirrel while searching for cheese and hiding from the owl. If, for example, the agent is consistently eaten by the owl every time it sees a spider web or a squirrel it will associate these observations with danger and fail to realize that they are correlated due to the presence of a tree. In more realistic scenarios, navigational problems become challenging because there are multiple observables that are correlated to task-relevant variables that do not hold the true cause. Such confounders are abundant in the real world and today's algorithms used for navigation face difficulties in generalizing to domain shifts by failing to account for the causal relationships between the agent and the environment.
In this project you will implement and study neural networks that are able to extract causal dynamics and structures and are trained either with supervised or reinforcement learning (Fig 1B). You will learn to apply tools from both physics, neuroscience and artificial intelligence to understand how the causal structure of the observed world is represented in these models, and how these representations impact problems found in navigation and can be reused on e.g. temporal recognition tasks.
The project includes collaborations with leading groups at Harvard University, University of Pennsylvania and the University of California San Diego.
Requirements
- You must have a master degree in physics, computational neuroscience, mathematics, statistics, or artificial intelligence.
- Documented experience from computational modeling, scientific programming or implementation and studies of neural network systems is an advantage.
Supervisors
Professor Anders Malthe-Sørenssen
Call 2b: Project start winter 2022/2023
This project is in call 2b, starting winter 2022/2023.