A research team led by Associate Professor Feng Weijia from the School of Computer and Information Engineering at Tianjin Normal University (TNU) has recently published a paper titled "Active Multimodal Distillation for Few-shot Action Recognition" at IJCAI 2025, a top-tier conference in artificial intelligence, which is recognized as a CCF-A class conference by the China Computer Federation.

Addressing the limitations of existing few-shot action recognition methods—such as their primary reliance on single-modal data, constrained performance on complex actions, and lack of dynamic evaluation of modality reliability—the paper proposes a novel active inference-based multimodal framework. This framework significantly enhances recognition performance through dynamic identification of sample-level reliable modalities, bidirectional knowledge distillation, and adaptive fusion. The proposed model first employs an active instance reasoning module to dynamically assess and select the most reliable modality for each query sample using variational free energy. Subsequently, an active mutual distillation module facilitates the bidirectional transfer of task knowledge from reliable modalities to unreliable ones, thereby strengthening their feature representations. Finally, an adaptive multimodal reasoning module generates the detection outcome by performing a weighted fusion based on the confidence level of each modality. Experimental results demonstrate that the proposed model achieves superior efficiency in utilizing multimodal information and significantly outperforms existing methods like STRM, TRX, and AFMAR in recognition accuracy across four benchmark datasets: Kinetics-400, SSv2, HMDB51, and UCF101.

Associate Professor Feng Weijia is the first author, with TNU listed as the primary institution. This research was supported by the General Programs of the National Natural Science Foundation of China (61602345, 62002263, 62302333), the National Key Research and Development Program of China (2019YFB2101900), the Enterprise R&D Special Project of Tiankai Higher Education Innovation Park (23YFZXYC00046), and the 2024 University-Industry Innovation Fund of the Chinese Ministry of Education (2024HY015).

Article Links: https://arxiv.org/pdf/2506.13322

By He Jierui