Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Continual Multimodal Contrastive Learning
Authors: Xiaohao Liu, Xiaobo Xia, See-Kiong Ng, Tat-Seng Chua
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Beyond our theoretical contributions, we conduct experiments on multiple datasets by comparing our method against advanced continual learning baselines. The empirical results further support our claims and demonstrate the efficacy of our method. |
| Researcher Affiliation | Academia | Xiaohao Liu Xiaobo Xia See-Kiong Ng Tat-Seng Chua National University of Singapore EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 DNS for Continual Multimodal Contrastive Learning (Training) ... Algorithm 2 DNS for Continual Multimodal Contrastive Learning (Inference) ... Pseudo-code for gradient updates |
| Open Source Code | Yes | Our codes are available at https://github.com/Xiaohao-Liu/CMCL. |
| Open Datasets | Yes | We evaluate on 7 multimodal datasets, including UCF101 [29], ESC50 [30], NYUDv2 [31], VGGSound-S [32]3, Clotho [33], TVL [34], and LLVIP [35]. ... All clips have been manually extracted from public field recordings available on Freesound.org. ... https://www.crcv.ucf.edu/data/UCF101.php |
| Dataset Splits | Yes | In our case, we select 70 examples for each category for training, and 10 examples for each category for testing to balance the distribution. ... To evaluate the performance, we randomly split the data with a ratio of 4:1 for training (1,600) and testing (400). ... In this version, 47,584 examples are used for training, and 654 examples are for testing. |
| Hardware Specification | Yes | All the experiments are conducted with 4 NVIDIA RTX A5000 GPUs and 256GB memory. |
| Software Dependencies | No | We utilize Py Torch [36] for implementation. The model is optimized via an Adam W optimizer [37] with a learning rate of 0.0001 and weight decay of 0.001. |
| Experiment Setup | Yes | The model is optimized via an Adam W optimizer [37] with a learning rate of 0.0001 and weight decay of 0.001. The batch size is set to 64. We approximate the null space projection with truncated SVD, where a minimum eigenvalue Ξ»min is set to 0.01 for Image Bind and Unibind, while 0.0001 for Language Bind. ... All models are trained for 5 epochs at every step. |