I am using computational neural network models to study visual perception. While the models and tasks to be solved are similar to those used in Machine Learning, the learning rules are constrained to be biologically plausible. Taking inspiration from the brain, the predictive coding framework describes visual perception as a generative process, in which top-down signals in the cortex reconstruct activity patterns in earlier areas. From this basis, my research diverges into the following branches:

1. Self-supervised learning of transformation-invariant representations: How can the brain simultaneously learn representations that are invariant to transformations (rotating, scaling, lateral shifts) AND a generative model to reconstruct partially occluded inputs from these representations?
2. Self-supervised representation learning on complex images (Master thesis project of Robin Weiler).
3. Implementation of predictive coding networks in dynamic, spiking neural networks (in collaboration with Kwangjun Lee).
4. Integration of predictive motor-feedback with visual perception (in collaboration with Giovanni Pezzulo).