Recovering Latent Causal Factor for Generalization to Distributional Shifts
Authors: Xinwei Sun, Botong Wu, Xiangyu Zheng, Chang Liu, Wei Chen, Tao Qin, Tie-Yan Liu
NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The utility of our approach is verified by improved generalization to distributional shifts on various real-world data. Our code is freely available at https://github.com/wubotong/La CIM. We first use simulated data to verify the correctness of the identifiability claim. Then, to demonstrate the utility, we test our approach on real-world data, consistently achieving better generalization to the new distribution; besides, we find that our inferred causal factor can be concentrated in highly explainable semantic regions for the task of image classification. Experimentally (in sec. 5.2), our approach generalizes better to distributional shifts, compared with others. 5 Experiments We first verify the identifiability claims of theorem 4.4 in sec. 5.1. Then we evaluate La CIM on real-world data in sec. 5.2: Non-I.I.D. Image dataset with Contexts (NICO); Colored MNIST (CMNIST); Alzheimer s Disease Neuroimaging Initiative (ADNI www.loni.ucla.edu/ADNI for early prediction of Alzheimer s Disease), to verify the generalization ability of our method on the target domain with distributional shifts. Table 1: Accuracy (%) on test domain. Average over 10 runs. |
| Researcher Affiliation | Collaboration | 1 Microsoft Research Asia, Beijing, 100080 2 Peking University, Beijing, 100871 |
| Pseudocode | No | No pseudocode or algorithm block is explicitly labeled or formatted in the paper. |
| Open Source Code | Yes | Our code is freely available at https://github.com/wubotong/La CIM. |
| Open Datasets | Yes | We verify the generalization ability of La CIM on three data: NICO, CMNIST and ADNI. NICO. We consider the cat/dog classification in Animal dataset in NICO, a benchmark for non-i.i.d problem in [20]. Colored MNIST (CMNIST). Alzheimer s Disease Neuroimaging Initiative (ADNI www.loni.ucla.edu/ADNI for early prediction of Alzheimer s Disease). |
| Dataset Splits | No | The paper explicitly states training and test domains, but does not provide specific details or percentages for a separate validation split or how it was used in the context of data partitioning. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used, such as GPU/CPU models or memory. |
| Software Dependencies | No | The paper mentions using SGD as optimizer but does not specify software names with version numbers for libraries like PyTorch, TensorFlow, or Python. |
| Experiment Setup | Yes | We implement SGD as optimizer, with learning rate (lr) 0.5 and weight decay (wd) 1e-5 for CMNIST; lr 0.01 with decaying 0.2 every 60 epochs, wd 5e-5 for NICO and ADNI (wd is 2e-4). The batch-size are set to 256, 30 and 4 for CMNIST, NICO, ADNI. |