Learning Semantically Meaningful Representations Through Embodiment
How do humans acquire a meaningful understanding of the world with little to no supervision or semantic labels provided by the environment? I investigate embodiment with a closed loop between action and perception as one key component in this process.
I try to apply the idea of embodied learning by training reinforcement learning agents in 3D virtual environments with high dimensional visual observations. This is done with little to no external supervision and learning is sometimes aided by curiosity. I investigate the kind of representations of the sensory input that are learned in the embodied agents, their sparseness, similarity to representations found in animals, their semantic content, robustness, generalizability as well as their similarities to representations found in supervised neural networks.
I further hypothesize that several of the common shortcomings of conventional ANNs such as their vulnerability to adversarial attacks and over-reliance on low-level features originate from the unnatural way in which they are trained. My goal is to show that embodied and curiosity driven exploration of the world leads to more robust and disentangled representations of objects and can even make close to zero-shot object labeling possible.
Clay, V., König, P., Pipa, G., & Kühnberger, K.-U. (2021). Fast concept mapping: The emergence of human abilities in artificial neural networks when learning embodied and self-supervised, .arXiv:2102.02153.
[Hackathon] Predicting the Spread of Disease
During the course of a 1 week hackathon organized by the Robert Koch Institute and the University of Osnabrück we developed a concept for using machine learning to make predictions about the spread of diseases. We also developed an interactive interface to view disease spread in Germany over time.
With thins project we won the first price at the hackathon and a trip to the Robert Koch Institute in Berlin.
Within a study project during my master I used artificial neural networks to make predictions based on a big data set of hospital patient data. For example the symptoms displayed at hospital admittion can be predictive of the diagnosis, length of stay and treatment.