Causal Inference over Visual-Semantic-Aligned Graph for Image Classification

Authors: Lei Meng, Xiangxian Li, Xiaoshuo Yan, Haokai Ma, Zhuang Qi, Wei Wu, Xiangxu Meng

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results from two benchmarking datasets show that VSCNet effectively builds the visual-semantic associations from images and leads to better performance than the state-of-the-art methods with enriched predictive information.
Researcher Affiliation Academia 1School of Software, Shandong University, Jinan, China 2Shandong Research Institute of Industrial Technology, Jinan, China 3School of Mechanical, Electrical & Information Engineering, Shandong University, Weihai, China EMAIL, EMAIL
Pseudocode No The paper describes the methodology with detailed mathematical formulations and module descriptions but does not include structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide any statement or link indicating that the source code for the methodology is openly available.
Open Datasets Yes To demonstrate the generality of VSCNet in different domains, the cross-modal image classification datasets Ingredient-101 (Bola nos, Ferr a, and Radeva 2017) and NUS-WIDE (Chua et al. 2009) are used in the experiments.
Dataset Splits Yes Datasets #Classes #Words #Training #Testing Ingredient-101 101 446 68,175 25,250 NUS-WIDE 81 1000 121,962 81,636
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or cloud computing instance specifications used for running the experiments.
Software Dependencies No The paper mentions using Adam optimizer, ResNet models (RN18, RN50), ViT, and ART clustering algorithm, but does not provide specific version numbers for these software components or the programming language/libraries used (e.g., Python, PyTorch).
Experiment Setup Yes During training, the batch size is fixed at 64 and the Adam optimizer is adapted with a learning rate chosen from 5e-5 to 5e-3. The decay rate of learning rate was chosen from 0.1 and 0.5 with a interval of 4 epochs. For VSCNet, since it is a model-agnostic framework, we investigated the Res Net18 (RN18) (He et al. 2016), Res Net50 (RN50) and Vi T (Dosovitskiy et al. 2021) as base models. We use ART (Meng, Tan, and Wunsch 2015; Meng, Tan, and Xu 2013; Wang et al. 2024b; Chen et al. 2023) as our clustering algorithm, and set the clustering parameter from 0.8 to 0.9. During the training of the CSNet, we set the upper bound factor as 0.2 and the lower bound factor as 0.8, respectively. The loss weights αl are chosen from {0.1, 1, 10, 100}. The weight of filtering αf is selected from 0.1 to 0.5. The top number J of regions is selected from {5, 10, 15}, the number M of regions is selected from {3, 4, 5}.