SciPost Submission Page
AdvNF: Reducing Mode Collapse in Conditional Normalising Flows using Adversarial Learning
by Vikas Kanaujia, Mathias S. Scheurer, Vipul Arora
Submission summary
| Authors (as registered SciPost users): | Vipul Arora |
| Submission information | |
|---|---|
| Preprint Link: | https://arxiv.org/abs/2401.15948v2 (pdf) |
| Date submitted: | 2024-04-12 10:18 |
| Submitted by: | Arora, Vipul |
| Submitted to: | SciPost Physics |
| Ontological classification | |
|---|---|
| Academic field: | Physics |
| Specialties: |
|
| Approaches: | Theoretical, Computational |
Abstract
Deep generative models complement Markov-chain-Monte-Carlo methods for efficiently sampling from high-dimensional distributions. Among these methods, explicit generators, such as Normalising Flows (NFs), in combination with the Metropolis Hastings algorithm have been extensively applied to get unbiased samples from target distributions. We systematically study central problems in conditional NFs, such as high variance, mode collapse and data efficiency. We propose adversarial training for NFs to ameliorate these problems. Experiments are conducted with low-dimensional synthetic datasets and XY spin models in two spatial dimensions.
Author comments upon resubmission
https://scipost.org/submissions/2401.15948v1/
List of changes
1. In Section Introduction, page no. 4, paragraph 2, line no. 20, we have added the advantages of our method as well as reply to the first point raised in Report 2 by the referee.
2. In Appendix C, under Hyperparameter Details, we have addressed the third point raised in Report 2 .
3. Figs. 5 & 6 have been updated along with captions as referred to in points 5, 6, and 7 raised in Report 2.
4.Typo above Eq. 7 has been amended as raised in point 10 in Report 2.
5. We have added the results for 32x32 lattice size trained on 10 temperature settings in the revised manuscript in Table 8 and Fig.7 .
6. In Conclusion, Paragraph 4, Page no. 18, line no 12, we have discussed the second point raised in Report 1.
7. In Conclusion, last paragraph, Page-19, we have included the future work addressing scaling up to large lattices with factorizable model.