Concept Bottleneck Generative Models
Authors: Aya Abdelsalam Ismail, Julius Adebayo, Hector Corrada Bravo, Stephen Ra, Kyunghyun Cho
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | On multiple datasets across different types of generative models, steering a generative model, with the CB layer, outperforms all baselines in some cases, it is 10 times more effective. In addition, we show how the CB layer can be used to interpret the output of the generative model and debug the model during or post training. 4 EXPERIMENTS & RESULTS |
| Researcher Affiliation | Collaboration | Aya Abdelsalam Ismail1,2 Julius Adebayo3 Héctor Corrada Bravo1 Stephen Ra1,2 Kyunghyun Cho1,2,4,5 1Genentech 2Prescient Design 3Guide Labs 4Department of Computer Science, New York University 5Center for Data Science, New York University |
| Pseudocode | No | The paper does not contain any pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/prescient-design/CBGM |
| Open Datasets | Yes | Datasets. We consider the following datasets: (a) The 64 × 64 version of Celeb Faces Attributes (Celeb-A), (Liu et al., 2015)...; (b) a curated subset of aesthetic LAION (Schuhmann et al., 2022) dataset...; (c) The Caltech-UCSD birds species (CUB) (Wah et al., 2011)...; (d) Flickr Faces-HQ (FFHQ) human faces dataset (Karras et al., 2019)...; and (e) the Color-MNIST dataset (Deng, 2012)... |
| Dataset Splits | No | The paper mentions tracking concept loss on a "validation set" and using a "held-out test set" for concept classifiers, but it does not specify concrete train/validation/test splits (e.g., percentages or sample counts) for the main generative models, which are crucial for reproducibility. |
| Hardware Specification | Yes | Overall, training required 240 V100s gpus hours. |
| Software Dependencies | No | The paper mentions various models and open-source implementations but does not list specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | No | The paper mentions hyperparameters α and β controlling loss importance but does not provide their specific values. It refers to training strategies from other papers without detailing them within the text, which is insufficient for reproduction without external lookup. |