Counterfactual Generation with Identifiability Guarantees
Authors: Hanqi Yan, Lingjing Kong, Lin Gui, Yuejie Chi, Eric Xing, Yulan He, Kun Zhang
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our theoretically grounded framework achieves state-of-the-art performance in unsupervised style transfer tasks, where neither paired data nor style labels are utilized, across four large-scale datasets. We validate our theoretical discoveries by demonstrating state-of-the-art performance on the unsupervised style transfer task, which demands representation disentanglement, an integral aspect of counterfactual generation. |
| Researcher Affiliation | Academia | 1University of Warwick,2 Carnegie Mellon University, 3King s College London, 4Mohamed Bin Zayed University of Artificial Intelligence |
| Pseudocode | No | The paper includes diagrams of its framework (e.g., Figure 3), but it does not provide formal pseudocode blocks or algorithms labeled as such. |
| Open Source Code | No | The paper provides a link to download datasets ('The datasets can be downloaded via https://github.com/cookielee77/DAST.'), but it does not state that the source code for their proposed model (MATTE) is made available or provide a link for it. |
| Open Datasets | Yes | We use four domain datasets to train our unsupervised model, i.e., Imdb, Yelp, Amazon and Yahoo, and follow the data split provided by Li et al. [2019]. The datasets can be downloaded via https://github.com/cookielee77/DAST. |
| Dataset Splits | Yes | Table 1: Dataset on four domains. Domains Train Dev Test... We use early stops if the validation reconstruction loss does not decrease for three epochs. |
| Hardware Specification | Yes | We used a machine with the following CPU specifications: AMD EPYC 7282 CPU. We use NVIDIA Ge Force RTX 3090 with 24GB GPU memory. |
| Software Dependencies | Yes | The models were implemented in Py Torch 2.0. and Python 3.9. |
| Experiment Setup | Yes | The VAE network is trained for a maximum of 25 epochs and a mini-batch size of 64 is used. We use early stops if the validation reconstruction loss does not decrease for three epochs. For the encoder, we use the Adam optimizer and the learning rate of 0.001. For the decoder, we use SGD with a learning rate of 0.1. For the content and style flow, we use Adam optimizer and the learning rate is 0.001. We set three different random seeds and report the average results. We have performed a grid search of λsparsity [1E-4,1E-3,1E-2], λpartial [3E-5,3E-3,3E-1] and λc-mask [1E-4,1E-3,1E-2]. The best configuration is [λsparisty, λpartial, λc-mask] = [1E-4,3E-3,1E-4]. |