Walking the Web of Concept-Class Relationships in Incrementally Trained Interpretable Models
Authors: Susmit Agrawal, Deepika Vemuri, Sri Siddarth Chakaravarthy P, Vineeth N. Balasubramanian
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experimentation, we show that our approach obtains state-of-the-art classification performance compared to other concept-based models, achieving over 2 the classification performance in some cases. We also study the ability of our model to perform interventions on concepts, and show that it can localize visual concepts in input images, providing post-hoc interpretations. |
| Researcher Affiliation | Academia | Indian Institute of Technology Hyderabad EMAIL, EMAIL |
| Pseudocode | No | The paper describes the methodology using textual descriptions and mathematical formulas, but does not include any explicitly labeled 'Pseudocode' or 'Algorithm' blocks. |
| Open Source Code | Yes | Code https://github.com/Susmit-A/Mu CIL Appendix https://susmit-a.github.io/misc/appendix.pdf |
| Open Datasets | Yes | We perform a comprehensive suite of experiments to study the performance of Mu CIL on well-known benchmarks: CIFAR-100, Image Net-100 (INet-100), and Cal Tech-UCSD Birds 200 (CUB200). |
| Dataset Splits | No | The paper mentions evaluating performance on a 'validation split' and refers to 'dataset details' in the appendix, but it does not provide specific percentages, sample counts, or explicit splitting methodology for the training, validation, and test sets within the main text. |
| Hardware Specification | No | The paper does not provide any specific details regarding the hardware (e.g., GPU models, CPU types, memory) used to conduct the experiments. |
| Software Dependencies | No | The paper discusses the use of a transformer architecture and models like GPT 3.5, but it does not explicitly list specific software dependencies (e.g., programming languages, libraries, frameworks) along with their version numbers required to reproduce the experiments. |
| Experiment Setup | Yes | In the CL setting, we study our performance over 5 and 10 experiences using concept-based methods in conjunction with three well-known CL algorithms: Experience Replay (ER) (Rebuffi et al. 2017), AGEM (Chaudhry et al. 2019), and DER++ (Buzzega et al. 2020), with a replay buffer size of 500 (we study other variations of buffer size in the Appendix). ... We empirically found λ1 = 5 and λ2 = 10 to give the best performance overall in terms of FAA, LA and grounding similarity, with LA = 0.7722 and cosine similarity 0.998. |