Probabilistic Concept Bottleneck Models
Authors: Eunji Kim, Dahuin Jung, Sangha Park, Siwon Kim, Sungroh Yoon
ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5. Experiments We evaluate Prob CBM with one synthetic dataset and two real-world datasets. With these datasets, we demonstrate how Prob CBM effectively models uncertainty under diverse visual contexts presented in raw images. Additionally, we conduct an analysis to examine the ambiguity induced by image transformations. |
| Researcher Affiliation | Academia | 1Department of Electrical and Computer Engineering, Seoul National University, Seoul, Korea 2Interdisciplinary Program in Artificial Intelligence, Seoul National University, Seoul, Korea. |
| Pseudocode | Yes | Algorithm 1 provides detailed steps of our training scheme for the class predictor, which is described in Sec. 4.3. |
| Open Source Code | Yes | Code is publicly available at https://github.com/ejkim47/prob-cbm. |
| Open Datasets | Yes | We create synthetic data using the MNIST dataset (Le Cun et al., 2010) which contains images of 10 digits. |
| Dataset Splits | Yes | We create synthetic data using the MNIST dataset (Le Cun et al., 2010) which consists of 60,000 images for training and 10,000 images for testing. |
| Hardware Specification | No | The paper mentions utilizing the "Naver Smart Machine Learning (NSML) platform" but does not specify any particular hardware components such as GPU or CPU models. |
| Software Dependencies | No | All experiments are implemented using Py Torch (Paszke et al., 2017). The specific version number for PyTorch is not provided, nor are other software dependencies with version numbers. |
| Experiment Setup | Yes | For the synthetic dataset, we set Dc and Dy as 16 and 32, respectively. For the real-world datasets, Dc and Dy are set as 16 and 128, respectively. Ns is set as 50. We initialize a and d, which are learnable parameters for scaling given by Eqs. 3 and 5, as 5 and 10, respectively. We use Adam P optimizer (Heo et al., 2021) with the cosine learning rate scheduler (Loshchilov & Hutter, 2017). The learning rate is set to 10 3 for the pretrained weights and 10 2 for the randomly initialized weights and learnable parameters. λKL is set as 5 10 5. |