Bayesian Cross-Modal Alignment Learning for Few-Shot Out-of-Distribution Generalization
Authors: Lin Zhu, Xinbing Wang, Chenghu Zhou, Nanyang Ye
AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Numerical experiments demonstrate that Bayes-CAL achieved state-of-the-art Oo D generalization performances on two-dimensional distribution shifts. |
| Researcher Affiliation | Academia | Lin Zhu, Xinbing Wang, Chenghu Zhou, Nanyang Ye* Shanghai Jiao Tong University, Shanghai, China zhulin sjtu@sjtu.edu.cn, xwang8@sjtu.edu.cn, zhouchsjtu@gmail.com, ynylincoln@sjtu.edu.cn |
| Pseudocode | No | The paper describes methods but does not include structured pseudocode or an algorithm block. |
| Open Source Code | Yes | Our code is available at https://github.com/Lin LLLL/Bayes CAL. |
| Open Datasets | Yes | We evaluate Bayes-CAL on datasets that cover both diversity shift and correlation shift: datasets dominated by correlation shift (NICO (He, Shen, and Cui 2021) and Colored Cats Dogs), and datasets dominated by diversity shift (PACS (Li et al. 2017) and VLCS (Torralba and Efros 2011)). |
| Dataset Splits | Yes | for each category, we randomly sample a 16-shot training set and a 16-shot validation set (an 8-shot training set and a 64-shot validation set for NICO due to the validation accuracy up to 100% if the validation set is too small) from each domain. and a 20-times random search for each of 3 pairs of weight initialization and training-validation data. and the threshold is selected by 95% correctly classified validation examples are detected into examples with high prediction confidence. |
| Hardware Specification | No | No specific hardware details like GPU/CPU models or memory specifications were provided for running experiments. |
| Software Dependencies | No | The paper mentions software components like CLIP, ResNet-50, and Word2Vector, but does not provide version numbers for general software dependencies (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | The number of context tokens is set as 16, the class token position (CTP) is a hyper-parameter that can be set as end or middle, and the class-specific context (CSC) can be set as True or False. The three additional hyper-parameters introduced by our paradigm are λ1, λ2, and λ3, corresponding to the coefficients of the three regularization terms. and For all experiments, we set the max epoch as 30 unless otherwise specified and Setting the hyper-parameters of Bayes-CAL as (0.1, 0, 0.1) and the max epoch as 50. |