Provably Improving Generalization of Few-shot models with Synthetic Data
Authors: Lan-Cuong Nguyen, Quan Nguyen-Tri, Bang Tran Khanh, Dung D. Le, Long Tran-Thanh, Khoat Than
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments results show that our approach demonstrates superior performance compared to state-of-the-art methods, outperforming them across multiple datasets. Section 5, titled 'Experiments', details the experimental settings, baselines, datasets, implementation details, main results, and ablation studies. Tables 1, 2, 3, 4, 5 and 6 present quantitative performance metrics on various datasets. |
| Researcher Affiliation | Collaboration | The authors are affiliated with: 1FPT Software AI Center, 2Hanoi University of Science and Technology, 3Vin University, 4University of Warwick. FPT Software AI Center is an industry affiliation, while Hanoi University of Science and Technology, VinUniversity, and University of Warwick are academic institutions. This mix indicates a collaborative affiliation. |
| Pseudocode | Yes | The paper includes 'Algorithm 1 Fine-tuning few-shot models with synthetic data' and 'Algorithm 2 Lightweight version', which are clearly labeled algorithm blocks detailing the proposed methods. |
| Open Source Code | No | The paper does not contain an explicit statement or a link to a repository for the open-source code of the methodology described. It mentions using the FAISS library and Stable Diffusion, which are third-party tools, but does not provide its own implementation code. |
| Open Datasets | Yes | The paper evaluates its method on 10 common datasets for few-shot image classification: FGVC Aircraft (Russakovsky et al., 2015), Caltech101 (Li et al., 2022), Food101 (Bossard et al., 2014), Euro SAT (Helber et al., 2019), Oxford Pets (Parkhi et al., 2012), DTD (Cimpoi et al., 2014), SUN397 (Xiao et al., 2010), Stanford Cars(Krause et al., 2013), and Flowers102 (Nilsback & Zisserman, 2008). These are all well-known, publicly available datasets, and the paper provides citations for them. |
| Dataset Splits | Yes | The paper specifies dataset usage for experiments: 'All experiments are conducted with 16 real shots and 500 synthetic images per-class, except our lightweight version, where only 64 synthetic images per class were utilized.' (Table 1 caption). It also mentions 'training/evaluation data split' in Section 5.1 when discussing DTD, implying specific data partitioning. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU models, CPU types, or memory amounts used for running its experiments. It only mentions fine-tuning 'CLIP Vi T-B/16 image encoder' and using 'Stable Diffusion (SD) version 2.1', but not the hardware these were run on. |
| Software Dependencies | Yes | The paper specifies 'Stable Diffusion (SD) (Rombach et al., 2022) version 2.1' as the generator used, which includes a specific version number for a key software component. |
| Experiment Setup | Yes | The paper provides specific experimental setup details, including: 'the guidance scale of SD is set to be 2.0', 'the number of clusters... typically around twice the number of classes', 'The hyperparameters to be tuned are: λ1, λ2... The values of λ1, λ2 vary between the data sets, but consistently maintain the ratio of 1/10... The hyperparameter λ was chosen at 4 for all datasets except Stanford Cars, where we set it at 1.' (Section 5.1). Additionally, Appendix B states: 'We train our models using Adam W... searching the learning rate in {2e 4, 1e 4, 1e 5, 1e 6} and the weight decay in {1e 3, 5e 4, 1e 4}', and 'We run the K-means clustering step for 300 iterations... For the classifier tuning phase, we train for 50 epochs for the full approach and 150 epochs for the lightweight approach'. |