TY  - CONF
AU  - Lagemann, Kai
AU  - Lagemann, Christian
AU  - Mukherjee, Sach
TI  - INVARIANCE-BASED LEARNING OF LATENT DYNAMICS
M1  - DZNE-2025-00172
SP  - 200372
PY  - 2024
AB  - We propose a new model class aimed at predicting dynamical trajectories from highdimensional empirical data. This is done by combining variational autoencodersand (spatio-)temporal transformers within a framework designed to enforce certainscientifically-motivated invariances. The models allow inference of system behavior at any continuous time and generalization well beyond the data distributionsseen during training. Furthermore, the models do not require an explicit neuralODE formulation, making them efficient and highly scalable in practice. We studybehavior through simple theoretical analyses and extensive empirical experiments.The latter investigate the ability to predict the trajectories of complicated systemsbased on finite data and show that the proposed approaches can outperform existingneural-dynamical models. We study also more general inductive bias in the contextof transfer to data obtained under entirely novel system interventions. Overall, ourresults provide a new framework for efficiently learning complicated dynamics in adata-driven manner, with potential applications in a wide range of fields includingphysics, biology, and engineering.
T2  - 12th International Conference on Learning Representations
CY  - 7 May 2024 - 11 May 2024, Vienna (Austria)
Y2  - 7 May 2024 - 11 May 2024
M2  - Vienna, Austria
LB  - PUB:(DE-HGF)8
UR  - https://pub.dzne.de/record/276091
ER  -