Noisy Correspondence Learning with Self-Reinforcing Errors Mitigation

Authors: Zhuohang Dang, Minnan Luo, Chengyou Jia, Guang Dai, Xiaojun Chang, Jingdong Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on challenging benchmarks affirm the efficacy and efficiency of SREM.
Researcher Affiliation Collaboration School of Computer Science and Technology, MOEKLINNS Laboratory, Xi an Jiaotong UniversitySGIT AI LabState Grid Corporation of ChinaUniversity of Technology SydneyBaidu IncMohamed bin Zayed University of Artificial Intelligence
Pseudocode No The paper does not include any explicitly labeled pseudocode or algorithm blocks.
Open Source Code No The paper does not provide an explicit statement or link for the open-sourcing of the code for the SREM methodology.
Open Datasets Yes We evaluate SREM using three image-text retrieval datasets, including COCO (Lin et al. 2014), Flickr30K (Young et al. 2014) and CC152K (Huang et al. 2021).
Dataset Splits Yes Following (Huang et al. 2021), we maintain 5K/5K and 5K/5K image-text pairs for validation/test, leaving the remainder for training.
Hardware Specification Yes We report the training overhead per epoch on CC152K using an NVIDIA Tesla A40 48G in Table 4.
Software Dependencies No The paper mentions using Adam optimizer but does not provide specific version numbers for software dependencies or libraries.
Experiment Setup Yes Specifically, we warm up the model for 5 epochs with Li2t c and Lt2i c to achieve initial convergence, followed by a 50 epochs training process. We employ a batch size of 128 and an Adam (Kingma and Ba 2014) optimizer with a learning rate of 2e-4 that will be decayed by 0.1 after 25 epochs.