Language Model as Visual Explainer

Authors: Xingyi Yang, Xinchao Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental To access the effectiveness of our approach, we introduce new benchmarks and conduct rigorous evaluations, demonstrating its plausibility, faithfulness, and stability. (Abstract) and 4 Experiment This section offers an in-depth exploration of our evaluation process for the proposed LVX framework...
Researcher Affiliation Academia Xingyi Yang Xinchao Wang National University of Singapore xyang@u.nus.edu, xinchao@nus.edu.sg
Pseudocode Yes C Pseudocode for LVX In this section, we present the pseudocode for the LVX framework, encompassing both the construction stage and the test stage. The algorithmic pipelines are outlined in Algorithm 1 and Algorithm 2.
Open Source Code Yes Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: Code has been uploaded as supplementary material. Data will be available. (Paper Checklist Q5)
Open Datasets Yes To address this, we developed annotations for three recognized benchmarks: CIFAR10, CIFAR100 [39], and Image Net [57], termed as H-CIFAR10, H-CIFAR100, and H-Image Net. (Section 4.1)
Dataset Splits Yes The model is trained on a labeled training set Dtr = {xj, yj}M j=1, and would be evaluated a test set Dts = {xj}L j=1. (Section 2) and We report the average score across all validation samples. (Section 4.1) and To address this, we developed annotations for three recognized benchmarks: CIFAR10, CIFAR100 [39], and Image Net [57]...
Hardware Specification No The paper does not specify the exact hardware (e.g., specific GPU or CPU models, memory, or cloud instance types) used for running the experiments.
Software Dependencies No The paper mentions software packages like PyTorch, torchvision, and timm but does not provide specific version numbers for these dependencies.
Experiment Setup Yes The model is optimized with SGD for 50 epochs on the training sample, with an initial learning rate in {0.001, 0.01, 0.03} and a momentum term of 0.9. The weighting factor is set to 0.1. (Section 4.1) and In our experiment, we performed five rounds of tree refinement. (Section 3.1)