Concept Gradient: Concept-based Interpretation Without Linear Assumption
Authors: Andrew Bai, Chih-Kuan Yeh, Neil Y.C. Lin, Pradeep Kumar Ravikumar, Cho-Jui Hsieh
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate empirically that CG outperforms CAV in evaluating concept importance on real world datasets and perform a case study on a medical dataset. |
| Researcher Affiliation | Academia | Andrew Bai Department of Computer Science University of California, Los Angeles andrewbai@cs.ucla.edu Chih-Kuan Yeh Department of Computer Science Carnegie Mellon University cjyeh@cs.cmu.edu Neil Y. C. Lin Department of Bioengineering University of California, Los Angeles neillin@g.ucla.edu Pradeep Ravikumar Department of Computer Science Carnegie Mellon University pradeepr@cs.cmu.edu Cho-Jui Hsieh Department of Computer Science University of California, Los Angeles chohsieh@cs.ucla.edu |
| Pseudocode | No | The paper describes the steps and definitions of Concept Gradients (CG) in text and mathematical formulas but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The code is available at github.com/jybai/concept-gradients. |
| Open Datasets | Yes | We experimented on CUB-200-2011 (Wah et al., 2011), a dataset for fine-grained bird image classification. ... We conducted the experiment on the Animals with Attributes 2 (Aw A2) dataset (Xian et al., 2018) |
| Dataset Splits | Yes | We combined all data together then performed a 80:20 split for the new training and validation set. ... The local attribution performance is measured by the average local recall@k (over samples) on a holdout testing set. The global attribution performance is measured by the average global recall@k (over classes) on a holdout testing set. |
| Hardware Specification | No | The paper does not specify any particular hardware components such as GPU models, CPU types, or memory used for conducting the experiments. |
| Software Dependencies | No | The paper does not provide specific version numbers for software dependencies or libraries (e.g., Python, PyTorch, TensorFlow, or other packages). |
| Experiment Setup | Yes | We searched for hyperparameters over a range of learning rates (0.01, 0.001), learning rate schedules (decaying by 0.1 for every 15, 20, 25 epochs until reaching 0.0001), and weight decay (0.0004, 0.00004). We optimized with the SGD optimizer. |