Interpolation Technique to Speed Up Gradients Propagation in Neural ODEs
Authors: Talgat Daulbaev, Alexandr Katrutsa, Larisa Markeeva, Julia Gusak, Andrzej Cichocki, Ivan Oseledets
NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | As a result, our method allows faster model training than the reverse dynamic method that was confirmed and validated by extensive numerical experiments for several standard benchmarks. |
| Researcher Affiliation | Academia | Talgat Daulbaev, Alexandr Katrutsa, Larisa Markeeva, Julia Gusak, Andrzej Cichocki, Ivan Oseledets Skolkovo Institute of Science and Technology Moscow, Russia |
| Pseudocode | No | The paper does not contain structured pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code of the proposed method can be found at Git Hub1. 1https://github.com/Daulbaev/IRDM |
| Open Datasets | Yes | We tested these methods on four toy datasets (2spirals, pinwheel, moons and circles) and tabular miniboone dataset [23]. |
| Dataset Splits | No | No explicit description of training/validation/test dataset splits, or specific information about a dedicated validation set, was found. While 'test loss' is mentioned in the context of stopping criteria, it's not explicitly defined as a validation set. |
| Hardware Specification | Yes | Every separate experiment is conducted on a single NVIDIA Tesla V100 GPU with 16Gb of memory [21]. |
| Software Dependencies | No | The paper mentions 'torchdiffeq' and 'DOPRI5' but does not provide specific version numbers for these software dependencies. |
| Experiment Setup | Yes | For toy datasets, we used the following hyperparameters: learning rate equals 10 3, the number of epochs was 10000, the batch size was 100, absolute and relative tolerances in the DOPRI5 solver were 10 5 and 10 5. |