Non-adversarial training of Neural SDEs with signature kernel scores
Authors: Zacharia Issa, Blanka Horvath, Maud Lemercier, Cristopher Salvi
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We perform experiments across five datasets. First is a univariate synthetic example, the benchmark Black-Scholes model... Finally we present a spatiotemporal generative example... We evaluate each training instance with a variety of metrics. The first is the Kolmogorov-Smirnov (KS) test... |
| Researcher Affiliation | Academia | 1Department of Mathematics, King s College London, London, United Kingdom. 2Department of Mathematics, Oxford University, Oxford, United Kingdom. 3The Oxford Man Institute, Oxford, United Kingdom. 4The Alan Turing Institute, London, United Kingdom. 5Department of Mathematics, Imperial College London, London, United Kingdom. |
| Pseudocode | No | The paper describes methods using mathematical equations and prose but does not include any explicit pseudocode blocks or algorithm listings. |
| Open Source Code | No | The paper does not include an explicit statement about releasing source code or a link to a code repository for the methodology described. |
| Open Datasets | Yes | Data is obtained from https://www.dukascopy.com/swiss/english/home/ (for foreign exchange data) and We fit our model on real LOB data from the NASDAQ public exchange [NMK+18]. |
| Dataset Splits | Yes | For all three discriminators, the training and test set were both comprised of 32768 paths and the batch size was chosen to be N = 128. |
| Hardware Specification | Yes | All experiments were run on a NVIDIA Ge Force RTX 3070 Ti GPU, except the experiment in Section 4.4 for which the NSPDE model was trained using a NVIDIA A100 40GB GPU. |
| Software Dependencies | No | The paper mentions 'torchsde package' and 'Py Torch' but does not specify their version numbers. |
| Experiment Setup | Yes | For all three discriminators, the training and test set were both comprised of 32768 paths and the batch size was chosen to be N = 128. We trained the SDE-GAN for 5000 steps, ϕsig for 4000 and ϕN sig for 10000 steps. (and) Learning rates were roughly proportional to the average size of the batched loss: as a rough guide, proportionality like ηG L(θ) 10 5 tended to yield good results, with the generator learning rate being around ηG 10 4 for the signature kernel(s), and ηD ηG 10 2 for the SDE-GAN. |