Unsupervised Causal Binary Concepts Discovery with VAE for Black-Box Model Explanation

Authors: Thien Q Tran, Kazuto Fukuchi, Youhei Akimoto, Jun Sakuma9614-9622

AAAI 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Finally, using multiple datasets, we demonstrate that the proposed method can discover useful concepts for explanation in this form. Experiment Experiment Setting In this section, we demonstrate our method using three datasets: EMNIST(Cohen et al. 2017), MNIST(Lecun et al. 1998) and Fashion-MNIST(Xiao, Rasul, and Vollgraf 2017). Quantitative Results We evaluate the causal influence of a concept set using the total transition effect (TTE), which is defined as [...] In Figure 8a, we show the test-time mutual information and the TTE values when the causal objective LCE uses the prior p (γ) (Eq. (9)), VAE model s prior p(γ) and when trained without LCE.
Researcher Affiliation Academia 1 University of Tsukuba 2 Riken AIP thientquang@mdl.cs.tsukuba.ac.jp, {fukuchi,akimoto,jun}@cs.tsukuba.ac.jp
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not contain an explicit statement about releasing source code or a link to a code repository.
Open Datasets Yes In this section, we demonstrate our method using three datasets: EMNIST(Cohen et al. 2017), MNIST(Lecun et al. 1998) and Fashion-MNIST(Xiao, Rasul, and Vollgraf 2017).
Dataset Splits No The paper mentions training on selected classes but does not provide specific train/validation/test split percentages, sample counts, or details on cross-validation setups needed for reproduction.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU specifications, or cloud computing instances used for running experiments.
Software Dependencies No The paper does not list specific software dependencies with version numbers (e.g., Python 3.x, TensorFlow 2.x, PyTorch 1.x).
Experiment Setup Yes The dimension of αi and β are K = 1 and L = 7, respectively. The explanation results of other datasets and further detailed experiment settings can be found in our extended version (Tran et al. 2021).