Stabilizing Backpropagation Through Time to Learn Complex Physics
Patrick Schnell, Nils Thuerey
Abstract
Of all the vector fields surrounding the minima of recurrent learning setups, the gradient field with its exploding and vanishing updates appears a poor choice for optimization, offering little beyond efficient computability. We seek to improve this suboptimal practice in the context of physics simulations, where backpropagating feedback through many unrolled time steps is considered crucial to acquiring temporally coherent behavior. The alternative vector field we propose follows from two principles: physics simulators, unlike neural networks, have a balanced gradient flow, and certain modifications to the backpropagation pass leave the positions of the original minima unchanged. As any modification of backpropagation decouples forward and backward pass, the rotation-free character of the gradient field is lost. Therefore, we discuss the negative implications of using such a rotational vector field for optimization and how to counteract them. Our final procedure is easily implementable via a sequence of gradient stopping and component-wise comparison operations, which do not negatively affect scalability. Our experiments on three control problems show that especially as we increase the complexity of each task, the unbalanced updates from the gradient can no longer provide the precise control signals necessary while our method still solves the tasks. Our code can be found at https://github.com/tum-pbs/StableBPTT .
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
1 top-tier paper has cited it since. Know what they found before your reviewers do.
Lune can draw on
the full paper · the top-tier papers that cite it · related work
Free to start. No credit card required.
Lineage
- Solver-in-the-Loop: Learning from Differentiable Physics to Interact with Iterative PDE-Solvers cites DiffTaichi: Differentiable Programming for Physical Simulation
- Accelerated Policy Learning with Parallel Differentiable Simulation cites DiffTaichi: Differentiable Programming for Physical Simulation
- Temporal Difference Learning: Why It Can Be Fast and How It Will Be Faster cites Solver-in-the-Loop: Learning from Differentiable Physics to Interact with Iterative PDE-Solvers
Cited by top-tier papers1
Builds on11
- DiffTaichi: Differentiable Programming for Physical SimulationYuanming Hu, Luke Anderson, Tzu-Mao Li, Qi Sun et al.ICLR 2020 · 479 citations
- Message Passing Neural PDE SolversJohannes Brandstetter, Daniel E. Worrall, Max WellingICLR 2022 · 410 citations
- Solver-in-the-Loop: Learning from Differentiable Physics to Interact with Iterative PDE-SolversKiwon Um, Robert Brand, Yun (Raymond) Fei, Philipp Holl et al.NeurIPS 2020 · 398 citations
- Learning to Control PDEs with Differentiable PhysicsPhilipp Holl, Nils Thuerey, Vladlen KoltunICLR 2020 · 221 citations
- Accelerated Policy Learning with Parallel Differentiable SimulationJie Xu, Viktor Makoviychuk, Yashraj Narang, Fabio Ramos et al.ICLR 2022 · 141 citations
Related papers
- Half-Inverse Gradients for Physical Deep LearningPatrick Schnell, Philipp Holl, Nils ThuereyICLR 2022 · 9 citations
- Coupled Oscillatory Recurrent Neural Network (coRNN): An accurate and (gradient) stable architecture for learning long time dependenciesT. Konstantin Rusch, Siddhartha MishraICLR 2021 · 121 citations
- Time-Efficient Reinforcement Learning with Stochastic Stateful PoliciesFiras Al-Hafez, Guoping Zhao, Jan Peters, Davide TateoICLR 2024 · 3 citations
- Learning long range dependencies through time reversal symmetry breakingGuillaume Pourcel, Maxence ErnoultNeurIPS 2025 · 9 citations
- Dynamics and Representation Structure of Local Approximations to Gradient-Based Learning in Linear Recurrent Neural NetworksEzekiel Williams, Alexandre Payeur, Guillaume LajoieICML 2026
