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..

SEGA: Shaping Semantic Geometry for Robust Hashing under Noisy Supervision

Authors: Yiyang Gu, Bohan Wu, Qinghua Ran, Rong-Cheng Tu, Xiao Luo, Zhiping Xiao, Wei Ju, Dacheng Tao, Ming Zhang

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on a range of widely-used retrieval datasets justify the superiority of our SEGA over extensive strong baselines under noisy supervision.
Researcher Affiliation Academia 1 State Key Laboratory for Multimedia Information Processing, School of Computer Science, PKU-Anker LLM Lab, Peking University 2 School of Mathematical Sciences, Peking University 3 Nanyang Technological University 4 UCLA 5 University of Washington EMAIL, EMAIL, EMAIL, EMAIL, EMAIL
Pseudocode Yes Algorithm 1: Training Procedure of SEGA
Open Source Code Yes The code can be found at https://github.com/dllab001/SEGA.
Open Datasets Yes We evaluate SEGA on four widely-used image retrieval benchmarks: CIFAR-10 [28], Flickr25K [24], NUS-WIDE [10], and MS COCO [36].
Dataset Splits Yes CIFAR-10 ... We sample 1,000 examples per class as queries. The remaining examples are utilized as the retrieval database, from which 500 images per class are further sampled for training. Flickr25k... We use 2,000 images as queries, and sample 10,000 images for training from the remaining retrieval set. NUS-WIDE... We employ 5,000 images as queries, and sample 5,000 images from the remaining retrieval set for training. MS COCO... use 5,000 images as queries. From the remaining retrieval set, 10,000 are sampled for training.
Hardware Specification Yes All experiments are implemented in Py Torch and run on a single NVIDIA A40 GPU in a standard Linux environment.
Software Dependencies No All experiments are implemented in Py Torch and run on a single NVIDIA A40 GPU in a standard Linux environment. (No version numbers provided for PyTorch or other specific libraries/software.)
Experiment Setup Yes We utilize stochastic gradient descent (SGD) with a momentum of 0.9 and a batch size of 24. The learning rate is initialized at 0.001, with weight decay set to 0.0004 and dropout rate to 0.5. The backbone network is initialized from a pretrained VGG-16 [49] model, consistent with all baseline. During mixup training, interpolation coefficients are sampled from a symmetric Beta distribution with α = 0.4 as in [66]. The percentile threshold qr for selecting clean samples is set to 0.3 by default.