Towards Understanding Extrapolation: a Causal Lens

Authors: Lingjing Kong, Guangyi Chen, Petar Stojanov, Haoxuan Li, Eric Xing, Kun Zhang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Through experiments on both synthetic and real-world data, we validate our theoretical findings and their practical implications.
Researcher Affiliation Academia 1 Carnegie Mellon University 2 Mohamed bin Zayed University of Artificial Intelligence 3 Broad Institute of MIT and Harvard, Cancer Program, Eric and Wendy Schmidt Center
Pseudocode No The paper does not contain any pseudocode or algorithm blocks.
Open Source Code Yes The code is provided here.
Open Datasets Yes We conduct experiments on Image Net-C [45] and Image Net100-C [46] with 15 different types of corruption. ...We conduct experiments on the CIFAR10-C, CIFAR100-C, and Image Net-C datasets [45]...
Dataset Splits No The paper mentions training and testing, but does not explicitly provide details about a validation dataset split (e.g., percentages or sample counts for validation).
Hardware Specification Yes The experiments are conducted with the Py Torch 1.11.0 framework, CUDA 12.0 with 4 NVIDIA A100 GPUs. ... The experiments are conducted with the Py Torch 1.13.0 framework, CUDA 11.7 with an NVIDIA A100 GPU.
Software Dependencies Yes The experiments are conducted with the Py Torch 1.11.0 framework, CUDA 12.0 with 4 NVIDIA A100 GPUs. ... The experiments are conducted with the Py Torch 1.13.0 framework, CUDA 11.7 with an NVIDIA A100 GPU.
Experiment Setup Yes We train all methods with Adam [68] and learning rate 2e 3 for 25 epochs. We fix the loss weights λcls = 1, λrecons = 0.1, λtgt_likelihood = 0.1, and λs_distance = 0.01 (for dense shifts) overall distance configurations. We only tune λKL from the interval {1e 1, 1e 2, 1e 3}. ... In the pre-train stage, we apply the Res Net50 [47] as the backbone network...