Extrapolation and learning equations
Authors: Georg Martius, Christoph H. Lampert
ICLR 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the ability of EQL to learn physically inspired models with good extrapolation quality by experiments on synthetic and real data. For this, we implemented the network training and evaluation procedure in python based on the theano framework Theano Development Team (2016). We will make the code for training and evaluation public after acceptance of the manuscript. |
| Researcher Affiliation | Academia | Georg Martius & Christoph H. Lampert IST Austria Am Campus 1, 3400 Klosterneuburg, Austria {gmartius,chl}@ist.ac.at |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. |
| Open Source Code | No | We will make the code for training and evaluation public after acceptance of the manuscript. |
| Open Datasets | Yes | For that we use recorded trajectories of a real double pendulum Schmidt & Lipson (2009). The data is taken from Deslattes et al. (2003) |
| Dataset Splits | Yes | The dataset contains two short trajectories. The first covers only part of the domain (input as well as output) and consists of 819 samples where 10% was used as validation set (randomly sampled) |
| Hardware Specification | No | The paper does not provide specific hardware details such as CPU, GPU models, or memory specifications used for experiments. |
| Software Dependencies | No | We implemented the network training and evaluation procedure in python based on the theano framework Theano Development Team (2016). |
| Experiment Setup | Yes | In all numerical experiments we use α = 0.001 and a mini-batch size of 20. [...] at the beginning of the training procedure (t < t1) we use no regularization (λ = 0), such that parameters can vary freely and reach reasonable starting points. Afterwards, we switch on the regularization by setting λ to a nonzero value, which has the effect that a sparse network structure emerges. Finally, for the last steps of the training (t > t2) we disable L1 regularization (λ = 0) but enforce the same L0 norm of the weights. This is achieved by keeping all weights w ∈ W 1...L that are close to 0 at 0, i. e. if |w| < 0.001 then w = 0 during the remaining epochs. [...] In practice, we use t1 = 1 4T and t2 = 19 20T, where T is total number of update steps. |