Research map / Continual learning / Online and replay-based

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

AAAI 2025 · Continual learning › Online and replay-based · approach: Dynamic expansion and isolation · 6 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ConSense, an exemplar-free class-incremental WiFi HAR framework that grows transformer MHSA prefixes per task and selectively retrains MLP neurons based on activation stability.

Problem

WiFi-based human activity recognition must learn new activities continually without forgetting old ones, but storing historical exemplars is impractical due to privacy and edge-device storage limits.

Approach

Combines dynamic model expansion (task-specific prefixes with a parallel adapter added to multi-head self-attention, with frozen previous prefixes) and selective retraining (freeze masks on MLP neurons whose average activations are stable across tasks) in a transformer with Gaussian range positional encoding, for exemplar-free class-incremental WiFi HAR.

Key results

• On WiAR long task (N=8), ConSense reaches 89.85% average accuracy versus 60.23% for the best baseline PRD, with 3.35M parameters.
• On MMFi short task (N=6), ConSense reaches 84.42% average accuracy, versus 54.46% for PRD.
• ConSense keeps average forgetting below 20% on all datasets (e.g., 12.89% on WiAR N=8), while using 1.50M to 3.35M parameters.

Measured effects reported in the paper

Surprising findings

• Selective retraining alone yields only modest gains over the no-strategy baseline on WiAR (59.37 vs 52.08 average accuracy), while dynamic expansion gives the larger gain.

Limitations

• Evaluated only on three public WiFi datasets with simulated class-incremental task splits
• Selective retraining contributes less than dynamic expansion in ablation
• Reported forgetting and accuracy depend on fixed hyperparameters such as the stability threshold epsilon

Experimental setup

Datasets: WiAR MMFi XRF

Benchmarks: WiAR MMFi XRF

Metrics: average accuracy average forgetting parameter count

Baselines: LWF PASS R-DFCIL PRD ConTraCon

Builds on: LwF PASS parallel adapter Gaussian range encoding ConTraCon

Closest papers by meaning

• CSI-Bench: A Large-Scale In-the-Wild Dataset for Multi-task WiFi Sensing (NeurIPS 2025 · similarity 0.87)
• WiFi CSI Based Temporal Activity Detection via Dual Pyramid Network (AAAI 2025 · similarity 0.87)
• HARDVS: Revisiting Human Activity Recognition with Dynamic Vision Sensors (AAAI 2024 · similarity 0.83)
• Scalable Hierarchical Deep Neural Network for Time Series Analysis in Wearable Sensor-based Human Activity Recognition (AAAI 2025 · similarity 0.82)
• MoPFormer: Motion-Primitive Transformer for Wearable-Sensor Activity Recognition (NeurIPS 2025 · similarity 0.82)
• Decomposing and Fusing Intra- and Inter-Sensor Spatio-Temporal Signal for Multi-Sensor Wearable Human Activity Recognition (AAAI 2025 · similarity 0.81)