Capturing Label Characteristics in VAEs
Authors: Tom Joy, Sebastian Schmon, Philip Torr, Siddharth N, Tom Rainforth
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 6 EXPERIMENTS Following our reasoning in 3 we now showcase the efficacy of CCVAE for the three broad aims of (a) intervention, (b) conditional generation and (c) classification for a variety of supervision rates, denoted by f. Specifically, we demonstrate that CCVAE is able to: encapsulate characteristics for each label in an isolated manner; introduce diversity in the conditional generations; permit a finer control on interventions; and match traditional metrics of baseline models. Furthermore, we demonstrate that no existing method is able to perform all of the above,2 highlighting its sophistication over existing methods. We compare against: M2 (Kingma et al., 2014); MVAE (Wu & Goodman, 2018); and our modified version of DIVA (Ilse et al., 2019). See Appendix C.4 for details. |
| Researcher Affiliation | Collaboration | Tom Joy1, Sebastian M. Schmon 1,2, Philip H. S. Torr1, N. Siddharth 1,3 & Tom Rainforth 1 1University of Oxford 2Improbable 3University of Edinburgh & The Alan Turing Institute |
| Pseudocode | No | No pseudocode or clearly labeled algorithm block was found. The paper describes methods using equations and prose. |
| Open Source Code | No | No explicit statement about releasing the source code for the methodology or a link to a code repository was found in the paper. |
| Open Datasets | Yes | To demonstrate the capture of label characteristics, we consider the multi-label setting and utilise the Chexpert (Irvin et al., 2019) and Celeb A (Liu et al., 2015) datasets. |
| Dataset Splits | No | No explicit details on specific training/validation/test dataset splits (e.g., percentages, sample counts, or explicit standard split citations) were found, only supervision rates for labeled data. |
| Hardware Specification | Yes | We trained the models on a Ge Force GTX Titan GPU. |
| Software Dependencies | No | No specific software dependencies with version numbers (e.g., libraries, frameworks, or operating systems) are mentioned in the paper. |
| Experiment Setup | Yes | Training consumed 2Gb for Celeb A and Chexpert, taking around 2 hours to complete 100 epochs respectively. Both models were optimized using Adam with a learning rate of 2e-4 for Celeb A respectively. |