Hamiltonian Generative Networks
Authors: Peter Toth, Danilo J. Rezende, Andrew Jaegle, Sébastien Racanière, Aleksandar Botev, Irina Higgins
ICLR 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this paper we introduce the first model that answers both of these questions without relying on restrictive domain assumptions. Our model, the Hamiltonian Generative Network (HGN), is a generative model that infers the abstract state from pixels and then unrolls the learned Hamiltonian following the Hamiltonian equations (Goldstein, 1980). We demonstrate that HGN is able to reliably learn the Hamil-tonian dynamics from noisy pixel observations on four simulated physical systems: a pendulum, a mass-spring and twoand threebody systems. Our approach outperforms HNN by a significant margin. |
| Researcher Affiliation | Industry | Peter Toth Deep Mind petertoth@google.com Danilo J. Rezende Deep Mind danilor@google.com Andrew Jaegle Deep Mind drewjaegle@google.com Sébastien Racanière Deep Mind sracaniere@google.com Aleksandar Botev Deep Mind botev@google.com Irina Higgins Deep Mind irinah@google.com |
| Pseudocode | No | The paper does not contain a clearly labeled 'Pseudocode' or 'Algorithm' block. While it describes integration methods, these are presented in prose and mathematical equations rather than structured algorithmic steps. |
| Open Source Code | Yes | More results and video evaluations are available at: http://tiny.cc/hgn |
| Open Datasets | No | The paper states, 'we generated four datasets analogous to the data used in Greydanus et al. (2019)' and describes the generation process, but it does not provide any specific link, DOI, or formal citation to make these generated datasets publicly available. |
| Dataset Splits | No | The paper specifies 'We generated 50 000 train and 10 000 test trajectories for each dataset' but does not explicitly mention or quantify a separate validation split or how it was used for reproduction. |
| Hardware Specification | No | The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory) used to run the experiments. It only mentions 'trained for around 16 hours' and 'using stochastic gradient descent' without specifying the machine. |
| Software Dependencies | No | The paper mentions using 'scipy (scipy.integrate.solve_ivp) (Jones et al., 2001)' but does not provide a specific version number for scipy or other key software libraries used for implementation (e.g., PyTorch, TensorFlow, or specific Python version). |
| Experiment Setup | Yes | We use Adam optimisier (Kingma & Ba, 2014) with learning rate 1.5e-4. ... We use Adam (Kingma & Ba, 2014) (learning rate 3e-4). ... The discrete timestep we use for the symplectic integrator update step is dt=0.125. ... The encoder network is a convolutional neural network with 8 layers, with 32 filters on the first layer, then 64 filters on each subsequent layer, while in the last layer we have 48 filters. ... The Hamiltonian transition network is a convolutional neural network of 6 layers, each consisting of 64 filters. ... The decoder network is a progressive network consisting of 3 residual blocks... |