Towards Compositionality in Concept Learning
Authors: Adam Stein, Aaditya Naik, Yinjun Wu, Mayur Naik, Eric Wong
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate CCE on five different datasets over image and text data. Our evaluation shows that CCE finds more compositional concept representations than baselines and yields better accuracy on four downstream classification tasks. |
| Researcher Affiliation | Academia | 1Department of Computer and Information Science, University of Pennsylvania, Pennsylvania, USA 2School of Computer Science, Peking University, Beijing, China. |
| Pseudocode | Yes | Algorithm 1 Compositional Concept Extraction Input: embeddings Z, num. attr. M, concepts per attr. K, subspace dimension S Initialize concepts C = {} for m = 1 . . . M do Initialize P Rd S such that P T P = I. Initialize K concepts V = {v1, . . . , v K}. repeat P = Learn Subspace(P, Z, V ) V = Learn Concepts(ZP, K) until Converged C = C V Z = Z ZPP T end for Return C |
| Open Source Code | Yes | Code and data are available at https://github.com/ adaminsky/compositional_concepts. |
| Open Datasets | Yes | CLEVR (Johnson et al., 2017) (vision), CUB (Wah et al., 2011) (vision), HAM10000 (Tschandl et al., 2018) (vision), Truth (Zou et al., 2023b) (language), and News (Mitchell, 1999) (language). |
| Dataset Splits | No | The paper references datasets and mentions using linear models for prediction and evaluation, which implies data splitting. However, it does not explicitly provide specific percentages, sample counts, or predefined split citations for train, validation, and test sets within the main body or directly linked appendices that I can parse without further deep inspection of external files. |
| Hardware Specification | No | The paper does not provide specific hardware details such as exact GPU/CPU models, processor types, or memory amounts used for running experiments. |
| Software Dependencies | No | The paper does not provide specific ancillary software details with version numbers, such as Python versions, PyTorch versions, or other library versions used in the experiments. |
| Experiment Setup | Yes | Table 16. Hyperparameters Dataset K M learning rate CLEVR 3 3 0.001 CUB 20 5 0.001 CUB-sub 5 4 0.1 Truth 12 10 0.001 Truth-sub [4, 2, 3] 3 0.001 HAM 20 25 0.02 News 15 30 0.001 |