Adaptive fusion label knowledge of prototype transduction propagation for few-shot learning

Purpose Deep learning has achieved remarkable results in many fields. It is data-driven and relies on large amounts of labeled data to train the models. However, data privacy and security can lead to high labeling costs in certain fields. The challenge of utilizing a minimal amount of target data to train more effective models has emerged as a novel research focus. Design/methodology/approach To address the problem of low accuracy in constructing class-level transduction propagation graphs with few samples, a prototype transduction propagation model that adaptively fuses class knowledge for few-shot learning is proposed. First, an improved convolutional network with fewer network parameters is proposed by combining feature reuse and multi-scale fusion to rich feature information. Second, an adaptive neighbor information aggregation mechanism is proposed to enrich the representation of the initial class prototype and control the fusion of two modal data. In addition, a cross-modal fusion mechanism is designed to fusion class knowledge and image class prototype representation to assist image recognition. Finally, the class-level transduction propagation graph is constructed by combining the intra-class similarity and inter-class difference to improve the model effect. Findings Following the analysis of the selected studies, we summarize the problems that existed during the research process and propose our approach. Further, experimental results show the effectiveness of our algorithm. The best accuracy of Adaptive fusion label knowledge of prototype transduction propagation for few-shot learning (ALPT) increases by 1.58% and 0.76% in the 5-way 5-shot and by 5.05% and 1.13% in the 5-way 1-shot on miniImagNet and tieredImageNet, respectively. Originality/value This paper analyzes the performance improvement of prototype transduction propagation for few-shot learning from various research directions. A comparative analysis of the experimental results on the current few-shot learning optimizations is performed to provide a reference for further research.