ConSense: Continually Sensing Human Activity with WiFi via Growing and Picking

TL;DR

ConSense tackles the challenge of continual WiFi-based HAR without storing past exemplars by applying a transformer with task-specific prefixes in the MHSA layer and a selective retraining scheme in the MLP. The model initializes with full training on the first task and then progressively expands with prefixes $P^t_K$ and $P^t_V$ for new tasks while keeping previous MHSA weights frozen, coupled with a neuron-activation–based freeze mask to preserve stable components. This exemplar-free design yields strong average accuracy gains across WiAR, MMFi, and XRF while reducing parameter count and maintaining forgetting below 20%, particularly excelling in long sequences and edge deployments. The approach demonstrates the importance of tailoring temporal-spatial CSI processing via Gaussian range encoding and selective parameter updates to achieve robust continual learning in dynamic wireless sensing scenarios.

Abstract

WiFi-based human activity recognition (HAR) holds significant application potential across various fields. To handle dynamic environments where new activities are continuously introduced, WiFi-based HAR systems must adapt by learning new concepts without forgetting previously learned ones. Furthermore, retaining knowledge from old activities by storing historical exemplar is impractical for WiFi-based HAR due to privacy concerns and limited storage capacity of edge devices. In this work, we propose ConSense, a lightweight and fast-adapted exemplar-free class incremental learning framework for WiFi-based HAR. The framework leverages the transformer architecture and involves dynamic model expansion and selective retraining to preserve previously learned knowledge while integrating new information. Specifically, during incremental sessions, small-scale trainable parameters that are trained specifically on the data of each task are added in the multi-head self-attention layer. In addition, a selective retraining strategy that dynamically adjusts the weights in multilayer perceptron based on the performance stability of neurons across tasks is used. Rather than training the entire model, the proposed strategies of dynamic model expansion and selective retraining reduce the overall computational load while balancing stability on previous tasks and plasticity on new tasks. Evaluation results on three public WiFi datasets demonstrate that ConSense not only outperforms several competitive approaches but also requires fewer parameters, highlighting its practical utility in class-incremental scenarios for HAR.

Figures (5)

• Figure 1: The initial model in the center is trained on two activities, walk and run. The retrained model on the left, fine-tuned with two new activities, fall and lie, loses the ability to recognize walk and run. In contrast, the retrained model on the right, which incorporates training data from both old and new activities, can recognize all four activities.
• Figure 2: Architecture of ConSense. Left part contains the framework. Right part details how the model dynamically expands and selectively retrains during continual learning from training session $1$ to training session $t$. As new tasks are introduced, the model dynamically expands with new prefixes in the MHSA layer. In the MLP, a selective retraining strategy is implemented to adjust neuron weights, preserving learned outcomes from stable neurons while updating unstable neurons to accommodate new tasks.
• Figure 3: Accuracy comparison of ConSense with other five methods on each task. The x-axis represents the $t$-th task, and the y-axis represents the accuracy of the $t$-th task, i.e., $A_t$.
• Figure 4: Ablation study of the impact of parallel adapter for long task on MMFi.
• Figure :