Recent advances in Time Series Foundation Models (TSFMs) have fundamentally revolutionized general time series analysis across domains like finance, retail, weather, and power. However, how to unlock the hidden capacity of general-purpose TSFMs for wearable activity recognition still remains largely unexplored, given severe sensor annotation scarcity and highly heterogeneous sensor data. To address these challenges, we propose DeepSenseMoE—a novel multi-scale convolution-based Mixture of Experts (MoE) module for parameter-efficient fine-tuning of general-purpose TSFMs to sensor-based activity recognition. DeepSenseMoE integrates three key innovations: (1) Multi-scale convolutional experts with different filter sizes responsible for capturing varying sensor contexts; (2) Shared-expert isolation mechanism compressing common activity knowledge into a single shared expert while reducing redundancy among routed experts; and (3) Hierarchical supervised contrastive alignment guiding experts to further learn discriminative activity features. Extensive experiments on three challenging HAR benchmarks demonstrate DeepSenseMoE's superiority, achieving up to 9.5% accuracy gains over state-of-the-art under few-shot and full-supervised settings, with only

Lightweight CNNs and Transformers have shown great promise in sensor-based human activity recognition (HAR), yet their structural synergies remain underexplored. This paper bridges this gap by integrating the MetaFormer paradigm—a general architecture abstracted from Transformers that structurally separates token mixing (i.e., self-attention) and channel mixing (i.e., feed-forward networks)—into efficient CNN design. While MetaFormer offers a powerful inductive bias, its standard self-attention mechanism is often computationally intensive for resource-constrained HAR. To address this, we revolutionize the classic MobileNetV3 architecture from a MetaFormer perspective, introducing Rep-MMB, a new family of pure lightweight CNNs. By leveraging structural reparameterization, Rep-MMB decouples multi-branch training-time complexity from efficient single-branch inference, enabling high accuracy with low latency. Evaluations on four public HAR benchmarks show that Rep-MMB outperforms state-of-the-art lightweight models in accuracy and efficiency, with practical validation on embedded devices. We hope that Rep-MMB may serve as a strong baseline to inspire future edge-deployed HAR research.

Human Activity Recognition (HAR) plays a pivotal role in ubiquitous computing. However, it remains constrained by the challenge of balancing fine-grained temporal modeling with real-time efficiency on resource-limited devices. While Transformer-based models excel at capturing long-range dependencies, they suffer from high computational costs, limiting their applicability on resource-constrained devices. Linear attention mechanisms improve efficiency but often discard negative signals and produce overly smooth, high-entropy attention distributions, impairing the extraction of fine-grained patterns and degrading classification accuracy in complex scenarios. In this work, we present ActiFormer, a novel sign-aware linear attention framework tailored for sensor-based HAR to overcome these limitations. To preserve bidirectional signal dynamics, we introduce Sign-Aware Attention, which explicitly models both same-sign and cross-sign interactions between queries and keys, effectively retaining negative signals crucial for accurate recognition. Furthermore, we propose a learnable entropy-scaling function that compensates for the exponential scaling effect lost in linear attention, originally provided by softmax, solving the high-entropy attention weight issue by amplifying the importance of critical temporal points. Extensive experiments on four benchmark HAR datasets demonstrate that ActiFormer consistently outperforms CNNs, standard Transformers, and state-of-the-art linear attention models, both in accuracy and efficiency. Its lightweight design supports real-time inference on edge devices such as the Raspberry Pi 5, highlighting its practical deployability in real-world applications.

Transformer models have demonstrated superior capability in capturing long-range temporal dependencies crucial for Sensor-Based Human Activity Recognition (HAR). However, the quadratic computational complexity inherent to the Softmax-Attention mechanism significantly impedes their deployment on resource-constrained wearable devices and real-time streaming tasks. To address this, we propose a novel Linear Transformer with Taylor Series Attention specifically tailored for the HAR domain, named TSA-Former. It leverages the first-order Taylor expansion to approximate the Softmax-Attention and utilizes the norm-preserving mapping to approximate the high-order non-linear information, resulting in a linear computational complexity. In addition, TSA-Former integrates a multi-branch architecture featuring multi-scale patch embedding, which enables the model to dynamically capture multi-scale temporal features while minimizing overhead. Experimental results across four public HAR benchmarks, namely UniMiB-SHAR, UCI-HAR, WISDM, and OPPORTUNITY, demonstrate that TSA-Former achieves state-of-the-art (SOTA) accuracy and efficiency, outperforming conventional Transformers and existing linear-attention models. Deployment experiments conducted on the Raspberry Pi 5 platform further validate the model’s superior low-latency and minimal power consumption profile, confirming its robust suitability for real-world embedded HAR applications. Code will be released.

Lightweight convolution-based neural networks (CNNs) are well suited for sensor-based human activity recognition (HAR) applications on resource-constrained edge devices with faster inference speed. However, the convolutional kernels are often limited to a small window range, which can only capture local details in time series sensor data, thus preventing further performance boost. Though Introducing self-attention into convolution can help to handle long-range dependence well, it might significantly slow down actual activity inference speed, due to high computational cost. In this paper, we introduce a new learning paradigm (star operation) and then present a lightweight Dual-Branch High-Order Interactions (DbHoi) block, which is computationally friendly for mobile HAR deployment. The proposed DbHoi block may implicitly transform raw sensor inputs into high-dimensional non-linear features, but actually operate in a low-dimensional feature space (analogs to the design principle of polynomial kernel tricks), without incurring extra computational overhead. Extensive experiments are conducted on three public HAR benchmarks including UCI-HAR, UniMiB-SHAR, and OPPORTUNITY, which demonstrate that our suggested DbHoi can consistently surpass various meticulously designed lightweight networks such as MobileNet, ShuffleNet, and GhostNet. Detailed ablation studies, visualizing representations, and on-device latency analyses further validate our insights with regards to the star operation, while underscoring its practical merit in real-world HAR deployment.

Inspired by recent success of foundation models in vision and language domains, time series foundation models (TSFMs) have garnered increasing attention in general time series analysis tasks like finance, weather, healthcare, and power. However, given high heterogeneity and severe annotation scarcity in time series sensor data, how to unlock the potential of large-scale general-purpose TSFMs for downstream activity recognition tasks remains yet unexplored? This paper makes the first attempt to address this timely challenge by adapting the self-supervised pre-trained TSFM (i.e., MOMENT) to few-shot activity recognition. We introduce a simple and efficient Sensor-Prompt Tuning (SPT) strategy, which employs multiple convolution-based sensor-friendly filters with a gating mechanism to act as learnable soft prompts, which can dynamically adapt sensor input space to the frozen TSFM backbone, effectively bridging domain gap between pre-training general time series data with wearable sensor stream. Extensive experiments across three public activity recognition benchmarks demonstrate that our SPT achieves up to 15.5% performance gains over existing state-of-the-art baselines under few-shot scenarios, while considerably outperforming other mainstream fine-tuning strategies with smaller than 1% of backbone parameters. Practical cloud-edge inference latencies are measured. This work offers a new prompt-tuning perspective on how to adapt pre-trained TSFMs for wearable activity recognition tasks. Code will be released.

During recent years, dynamic early-exit has provided a promising paradigm to improve the computational efficiency of deep neural networks by constructing multiple classifiers to let easy samples exit at shallow layers while avoiding redundant computations at deep exits, which has been seldom explored in the context of latency-aware human activity recognition (HAR) deployed on wearable devices. Particularly, most existing early-exit strategies have always treated all activity samples equally at each exit during training, which ignore such dynamic early-exit behavior at test-time, causing a potential mismatch between training and test. Intuitively, easy activity samples that often exit earlier at test-time should place more emphasis on the training loss of shallow classifiers, while hard activity samples should contribute more to the training loss of deep classifiers. To bridge this gap, this paper introduces a sample-reweighting approach for efficient activity inference, which employs a weight-predicting network to reweight the training loss of different activity samples at every exit. From a perspective of meta learning, a new optimization objective function is designed to jointly optimize both weight-predicting network and backbone network. We perform extensive experiments on three popular HAR benchmarks including UCI-HAR, WISDM, and UniMiB-SHAR, which demonstrate that while incorporating such test-time early-exit behavior into conventional training pipeline, it can consistently improve the accuracy-efficiency trade-offs under budgeted batch classification and anytime prediction patterns. Moreover, our approach has a natural advantage in handing class-imbalance HAR problem. Detailed ablation studies, visualized illustrations, and real hardware deployment are provided to support our statement.