Multifactor Sequential Disentanglement via Structured Koopman Autoencoders
Authors: Nimrod Berman, Ilan Naiman, Omri Azencot
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate our method extensively on two factor standard benchmark tasks where we significantly improve over competing unsupervised approaches, and we perform competitively in comparison to weaklyand self-supervised state-of-the-art approaches. The results show that our approach outperforms baseline methods in various quantitative metrics and computational resources aspects. |
| Researcher Affiliation | Academia | Nimrod Berman , Ilan Naiman , Omri Azencot Department of Computer Science Ben-Gurion University of the Negev {bermann,naimani}@post.bgu.ac.il, azencot@cs.bgu.ac.il |
| Pseudocode | No | The paper describes the architecture and process flow in text and diagrams (Figure 1) and provides architecture details in Table 5, but it does not include explicit pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at Git Hub. |
| Open Datasets | Yes | Sprites. Reed et al. (2015) introduced a dataset of animated cartoon characters. MUG. Aifanti et al. (2010) share a facial expression dataset TIMIT. Garofolo et al. (1992) made TIMIT available |
| Dataset Splits | Yes | Sprites. We use 9000 samples for training and 2664 samples for testing. MUG. Finally, we split the dataset such that 75% of it is used for the train set, and 25% for the test set. |
| Hardware Specification | No | The paper does not specify any particular hardware components such as specific GPU or CPU models used for running the experiments. |
| Software Dependencies | Yes | Our models are implemented in the Py Torch (Paszke et al., 2019) framework. |
| Experiment Setup | Yes | Regarding hyper-parameters, in our experiments, k is tuned between 40 and 200 and λrec, λpred and λeig are tuned over {1, 3, 5, 10, 15, 20}. ks is tuned between 4 and 20, and the ε threshold for the dynamic loss is tuned over {0.4, 0.5, 0.55, 0.6, 0.65}. The hyper-parameters are chosen through standard grid search. Table 6: Hyperparameter details. Dataset b k h #epochs λrec λpred λeig ks ϵ Sprites 32 40 40 800 15 1 1 8 0.5 MUG 16 40 100 1000 20 1 1 5 0.5 TIMIT 30 165 400 15 3 1 15 0 |