*Result*: AMPL: An adaptive meta-prompt learner for few-shot image classification.
*Further Information*
*Few-shot image classification is a challenging task that focuses on recognizing novel classes with only a few labelled images. While prompt-based learning has achieved remarkable success in natural language understanding with the advent of foundation models, its effectiveness in standard few-shot image classification has not been thoroughly explored. A straightforward way of integrating it with few-shot image classification might involve learning a specific prompt for each individual task, as commonly seen in existing prompt-based learning methods. However, given the diversity of few-shot tasks in practice, this method would significantly constrain the adaptability of the learned model and incur high computational costs due to the large number of individual few-shot learning tasks. To address this issue, this paper introduces a novel framework called Adaptive Meta-Prompt Learner (AMPL), which adaptively learns Meta-Visual-Prompts for diverse few-shot tasks. The meta-prompt learner leverages image patch features to generate visual prompts and can rapidly adapt to new tasks. Furthermore, to fully explore the potential of the proposed AMPL, this work designs a Token-Awareness Enhancement Module that utilizes the mutual awareness among tokens of the same class. This module helps the model capture task-aware and vision-sensitive concepts, thereby improving its robustness across various tasks and scenarios. Extensive experimental studies demonstrate that our work establishes new state-of-the-art classification performance across seven few-shot benchmark datasets. Notably, on the FC100 dataset, we achieve absolute improvements of 3.88 % and 7.96 % on classification accuracy compared with the hand-tuned prompt method, in the 1-shot and 5-shot scenarios, respectively. The repository is available at: https://github.com/ZhipingWoods/AMPL-main.
(Copyright © 2025 Elsevier Ltd. All rights reserved.)*
*Declaration of competing interest The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.*