Deep Energy-based Modeling of Discrete-Time Physics

Authors: Takashi Matsubara, Ai Ishikawa, Takaharu Yaguchi

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 4 Learning of Partial and Ordinary Differential Equations Comparative Models. We examined the proposed DGNet and comparative methods.
Researcher Affiliation Academia Takashi Matsubara Osaka University Osaka, Japan 560 8531 matsubara@sys.es.osaka-u.ac.jp Ai Ishikawa Kobe University Kobe, Japan 657 8501 a-ishikawa@stu.kobe-u.ac.jp Takaharu Yaguchi Kobe University Kobe, Japan 657 8501 yaguchi@pearl.kobe-u.ac.jp
Pseudocode Yes For reference, we introduce the case with a neural network that is composed of a chain of functions in Algorithm 1 in Appendix F.
Open Source Code Yes The proposed algorithm can be implemented in a similar way to the current automatic differentiation algorithm [20]; we provide it as a Py Torch library [31]1. 1https://github.com/tksmatsubara/discrete-autograd
Open Datasets Yes We employed Hamiltonian systems that were examined in the original study of the HNN [19], namely a mass-spring system, a pendulum system, and a 2-body system... We evaluated the models on the real pendulum dataset that were obtained by Schmidt and Lipson [39] following the study on the HNN [19]... We simulated the equation with a time step size of t = 0.001 for 500 steps and obtained 100 time series (90 for training and 10 for the test).
Dataset Splits No For the Kd V equation and Cahn Hilliard equation datasets, the paper states 'obtained 100 time series (90 for training and 10 for the test)' without mentioning a validation split. For other datasets, no explicit split information is provided beyond being used for training and test.
Hardware Specification No The paper does not provide specific hardware details (e.g., GPU/CPU models, memory) used for running its experiments. It only discusses the computational cost qualitatively.
Software Dependencies No The paper mentions implementation as a 'Py Torch library' and uses the 'Adam optimizer', but it does not specify explicit version numbers for these software dependencies, only providing a citation for PyTorch.
Experiment Setup Yes Each network was trained using the Adam optimizer [25] with a batch size of 200 and a learning rate of 0.001 for 10,000 iterations... We employed a neural network composed of a 1-dimensional convolution layer followed by two fully-connected layers. A convolution layer with a kernel size of 3 is enough to learn the central difference. The matrix G = D was implemented as a 1-dimensional convolution layer with the kernel of ( 1/2 x, 0, 1/2 x) and periodic padding. Following the study on the HNN [19], the activation function was the hyperbolic tangent, the number of hidden channels was 200, and each weight matrix was initialized as a random orthogonal matrix.