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..
Unleashing Foundation Vision Models: Adaptive Transfer for Diverse Data-Limited Scientific Domains
Authors: Qiankun Li, Feng He, Huabao Chen, Xin Ning, Kun Wang, Zengfu Wang
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Through extensive experiments on 10 datasets spanning domains such as generic, multimedia, biological, medical, industrial, agricultural, environmental, geographical, materials science, out-of-distribution (OOD), and 3D analysis, CLAdapter achieves state-of-the-art performance across diverse data-limited scientific domains, demonstrating its effectiveness in unleashing the potential of foundation vision models via adaptive transfer. |
| Researcher Affiliation | Academia | 1University of Science and Technology of China 2Ann Lab, Institute of Semiconductors, Chinese Academy of Sciences 3Nanyang Technological University |
| Pseudocode | No | The paper describes the methodology using mathematical equations and textual explanations, for example, in Section 3.2 'Cluster Attention Adapter (CLAdapter)' and Section 3.3 'Fine-tuning Strategy', but it does not present any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code is available at https://github.com/qklee-lz/CLAdapter. |
| Open Datasets | Yes | Pre-training Dataset and Backbones. In the era of big data, popular publicly available largescale 2D datasets include the Image Net-21K classification dataset [58] at the ten-million level, the LAION-400M image-text dataset [62] at the hundred-million level, and the LAION-2B image-text dataset [61] at the billion level. For the 3D domain, we utilize the Kinetics-400 [36], a large-scale dataset commonly used for action recognition, comprising approximately 260K video clips. Downstream Tasks. We experiment on 10 benchmarks across a broad spectrum of domains, including generic (Tiny-Image Net [38]), multimedia (UCF101 [66] and HDMB51 [37]), industrial (Ins PLAD-fault [78]), biological&medical (Break His [6] and HCRF [68]), agricultural (Apple Foliar Disease [72]), environmental&geographical (WHU-RS19 [81]), materials science (KTHTIPS-2b [51]), OOD (PACS [40]), and 3D analysis (above video) to demonstrate the versatility and effectiveness of our CLAdapter. |
| Dataset Splits | Yes | Tiny-Image Net. ... In the experiment, the data division adheres to official standards. PACS. ... the remaining data are divided into training and validation sets in a 5-fold manner, with a ratio of 1:4. Apple Foliar Disease. The data split method follows the previous work as training and validation sets with a 3:1 ratio. Break His. The dataset is split into training, validation, and testing sets in a 3:1:1 ratio, which is the same as the previous study [11]. HCRF. ...the HCRF dataset is divided into training, validation, and test sets as a 1:1:2 random stratified ratio. UCF101. The data division method follows the official release. |
| Hardware Specification | Yes | Our experimental setup is powered by four Nvidia Ge Force RTX 3090 GPUs, boasting 24 GB of memory, under the Ubuntu 20.04 environment. |
| Software Dependencies | Yes | Python 3.8.3 is chosen as the programming language, with the Py Torch 1.13.1 framework being utilized for model development. |
| Experiment Setup | Yes | Specifically, we adopt an input resolution of 224 224 pixels across all experiments. The learning rate is initialized at 1e 4, and models are trained for up to 100 epochs with a batch size of 16. We employ the Adam W optimizer, configuring it with momentum β1 = 0.9 and a weight decay of 1e 3, to adapt to the unique challenges presented by these varied datasets. The determination of cluster centers K for CLAdapter is fixed at 20 |