Zero-Parameter Attention Sharing Transformer for Joint Human Activity and Identity Recognition

Le résumé fourni par la source

WiFi-based human sensing is gaining popularity thanks to it not requiring additional devices and it not being as intrusive as cameras. Specifically, human features can be extracted from WiFi Channel State Information (CSI) to recognize human activities, identities, etc. However, most previous works rely on single-task learning models for recognition (e.g., to either recognize activities OR identities solely). The lack of cross-task knowledge sharing restricts these models to task-specific features and poor generalization. Recent studies have applied multi-task learning (MTL) to tackle this, but their cross-task sharing modules add vast amounts of extra parameters. Such massive parameters increase model complexity and reduce time efficiency. In this paper, we propose a novelZero-parameter Attention Sharing Transformer(ZAST) to efficiently recognize both activities and identities. In ZAST, aCross-task Attention on Attention(CAoA) mechanism computes the relevance of attention scores for cross-task knowledge sharing, as a new paradigm for lightweight MTL. To mitigate the perturbation caused by attention sharing, we formulate aMulti-head Similarity Loss(L-MS) for stable model training. We further equip ZAST withChannel-wise Squeeze and Excitation(CSE) that efficiently learns the channel correlations of CSI. Extensive experiments on four public datasets indicate that ZAST achieves state-of-the-art recognition performance with the lowest complexity and the highest efficiency.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.