Half-Inverse Gradients for Physical Deep Learning
Authors: Patrick Schnell, Philipp Holl, Nils Thuerey
ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method on three physical systems: controlling nonlinear oscillators, the Poisson problem, and the quantum dipole problem. |
| Researcher Affiliation | Academia | Patrick Schnell, Philipp Holl and Nils Thuerey Department of Informatics Technical University of Munich Boltzmannstr. 3, 85748 Garching, Germany {patrick.schnell,philipp.holl,nils.thuerey}@tum.de |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our code for the experiments presented in this paper is publicly available at https://github. com/tum-pbs/half-inverse-gradients. |
| Open Datasets | No | The datasets used in the experiments (nonlinear oscillators, Poisson problem, quantum dipole) are custom generated or sampled on-the-fly as described in sections B.2, B.3, and B.4. No public datasets or links to them are provided. |
| Dataset Splits | No | The paper mentions 'training data set' and 'test set' for some experiments (e.g., nonlinear oscillators, quantum dipole) but does not explicitly provide details for a validation split needed for reproduction. |
| Hardware Specification | Yes | Runtimes for the non-linear chain and quantum dipole were measured on a machine with Intel Xeon 6240 CPUs and NVIDIA Ge Force RTX 2080 Ti GPUs. The Poisson experiments used an Intel Xeon W-2235 CPU with NVIDIA Quadro RTX 8000 GPU. |
| Software Dependencies | Yes | As deep learning API we used Tensor Flow version 2.5. |
| Experiment Setup | Yes | Our method depends on several hyperparameters. First, we need a suitable choice of the learning rate. The normalizing effects of HIGs allow for larger learning rates than commonly used gradient descent variants. We are able to use η = 1 for many of our experiments. Second, the batch size b affects the number of data points included in the half-inversion process. |