What we do
We are an interdisciplinary research group focusing on principles of self-organization in neural systems, ranging from artificial neural networks to the hippocampus. By bringing together machine learning and computational neuroscience we explore ways of extracting representations from data that are useful for goal-directed learning, such as reinforcement learning.
Unsupervised Learning with Slow Feature Analysis
The goal of unsupervised learning algorithms is to learn a structure in data. This has the benefit of massively decreasing complexities of data-driven problems which would otherwise not be feasible to solve. A learning objective needs to be determined to guide this search of structure in the data. Our group is primarily interested in the learning algorithms defined by the slowness meta-learning objective: the most interesting things are the most slowly changing over time. A concrete method of realizing this is with the slow feature analysis algorithm, which extracts a set number of slowly changing features from a data set. As an example, this can identify from a wildlife video the location and angle of a present fish while discarding unimportant information, such as the value of every pixel at each time-frame.
Memory-based Reinforcement Learning
In recent years, we have seen machines becoming capable of solving tasks of increasing complexity without explicit instruction. For example, a single neural network outperformed several human experts in a selection of classical video games (Atari 2600) based on visual input. Interactive learning problems - an agent interacting with an environment - can be cast into the theoretical framework of "Reinforcement Learning", which is concerned with ways of maximizing accumulated rewards from such a setting. Our main interest lies in finding optimal behavior in visually complex environments, in which incompleteness of information is a common issue. One focus of our research lies in the design of algorithms that can use memory to overcome this problem by aggregating partial information from a history of observations.
System Level Modeling of the Hippocampus
The hippocampus is a central processing structure in the mammalian brain. In humans it is tightly linked to episodic memory; without it we would not be able to acquire new memories. In rodents the hippocampus has been found to house a neural encoding of the surrounding space in the form of place cells, grid cells, and head direction cells.
In our group we develop computational models for both of these aspects of the hippocampus in an effort to learn how these two fundamental functions interact with each other as well as how the underlying neuronal processes are capable of performing both of them in the same structure.