Deep neural networks (DNNs) have repeatedly proven to be successful in modelling complex functions. However, they are known to be data-hungry; as a result, they are often difficult to apply in scenarios where large datasets are not available. Additionally, DNNs are prone to break world governing laws (e.g. physics) in areas like pose estimation. This may be due to the model-free nature of DNNs. Having to learn how the world works from scratch may result in data hungriness and an incorrect world model. My goal is to construct model-based DNNs to obtain greater data efficiency and robustness. Since probabilistic modelling provides a language for describing structural relationships, I intend to utilise it to achieve model-based DNNs. In particular, my goal is to apply these model-based methods to 3D pose estimation for freely moving animals.
Publications
Lab members are shown in this color.
2024
Pawel A. Pierzchlewicz, Caio da Silva, James Cotton, Fabian H. Sinz
Platypose: Calibrated Zero-Shot Multi-Hypothesis 3D Human Motion Estimation
arXiv
|
2023
Pawel A. Pierzchlewicz, Konstantin F. Willeke, Arne F. Nix, Pavithra Elumalai, Kelli Restivo, Tori Shinn, Cate Nealley, Gabrielle Rodriguez, Saumil Patel, Katrin Franke, Andreas S. Tolias, Fabian H. Sinz
Energy Guided Diffusion for Generating Neurally Exciting Images
NeurIPS
|
Paweł A. Pierzchlewicz, Mohammad Bashiri, R. James Cotton, Fabian H. Sinz
Optimizing MPJPE promotes miscalibration in multi-hypothesis human pose lifting
ICLR - tiny paper
|
2022
Paweł A. Pierzchlewicz, R. James Cotton, Mohammad Bashiri, Fabian H. Sinz
Multi-hypothesis 3D human pose estimation metrics favor miscalibrated distributions
arXiv
|
Jiakun Fu, Konstantin F. Willeke, Pawel A. Pierzchlewicz, Taliah Muhammad, George H. Denfield, Fabian H. Sinz, Andreas S. Tolias
Hetereogenous orientation tuning across sub-regions of receptive fields of V1 neurons in mice
Cell Press Sneak Preview
|