Neural-HAR: Neural Activity Recognition

• Neural-HAR is a framework for neural-network–based human activity recognition that fuses multimodal sensor data—from inertial to radar—to detect activities in real time.
• It leverages diverse architectures such as CNNs, RNNs, Transformers, and spiking networks to address challenges across wearable, RF, and industrial sensing regimes.
• Neural-HAR advances practical deployment through efficient, low-power models and hardware co-design while tackling issues like label scarcity, class imbalance, and sensor variability.

Neural-HAR is best understood as neural-network-based human activity recognition: the use of learned representations and end-to-end or hybrid neural pipelines to infer activities from wearable inertial streams, motion-capture trajectories, RF/CSI/radar measurements, and multimodal sensor combinations. In the literature summarized here, the term spans CNN-, RNN-, Transformer-, spiking-, and weightless-network approaches, and it also appears as the title of a specific radar-oriented system, “Neural-HAR: A Dimension-Gated CNN Accelerator for Real-Time Radar Human Activity Recognition” (Wu et al., 26 Oct 2025).

1. Modalities, sensing regimes, and task scope

Neural-HAR has developed across several sensing regimes rather than a single canonical input type. A large part of the literature studies wearable inertial sensing from accelerometers and gyroscopes, often on smartphones, smartwatches, or wrist-worn devices, with fixed-window classification of activities such as walking, sitting, standing, lying, stairs, exercise, or gestures. Representative formulations include multimodal inertial fusion on UCL HAR and PAMAP2, sensor-wise Transformer modeling over 6-axis IMU windows, and self-supervised representation learning from month-scale wrist accelerometer streams later transferred to benchmark ADL datasets (Kasnesis et al., 2018, Ek et al., 2022, Sridhar et al., 2021).

A second line of work extends Neural-HAR beyond on-body sensing. RF-based and radar-based systems treat activity recognition as classification from reflected or perturbed electromagnetic signals rather than direct body-mounted kinematics. HAR-SAnet uses UWB signals and jointly learns from time-domain and frequency-domain RF representations (Chen et al., 2021). A CSI-based thesis treats Wi-Fi Channel State Information as a passive-radar-like modality, pairing spiking perception with symbolic reasoning (Bresciani et al., 2024). “Neural-HAR” in the narrow title sense denotes a compact radar micro-Doppler classifier, GateCNN, designed for real-time FMCW radar HAR on constrained FPGA hardware (Wu et al., 26 Oct 2025). This breadth suggests that Neural-HAR is not defined by one sensor family, but by the use of neural representation learning for activity inference under heterogeneous sensing physics.

A third regime is structured multimodality. In warehouse-process HAR, context-aware activity inference is built from MoCap-derived movement attributes plus process-step information (Lüdtke et al., 2021). In physiotherapy-oriented multimodal HAR, accelerometers, depth cameras, and pressure mats are fused with graph attention, mixture-of-experts routing, and a customized T5 encoder within a federated framework (Bandyopadhyay et al., 3 Aug 2025). The resulting scope of Neural-HAR is therefore wider than wearable IMU classification alone: it includes device-free sensing, industrial process context, and multimodal fusion under decentralized training.

2. Architectural patterns and representational design

Early Neural-HAR work already treated architectural bias as a first-order issue. PerceptionNet argued against immediate channel fusion and instead learned higher-level temporal features before cross-sensor interaction through a late $3 \times 15$ convolution, improving over early-fusion CNNs and LSTM-based baselines on multimodal inertial HAR (Kasnesis et al., 2018). ARN then introduced a dual-window residual design in which a short-window path captures local “spatial” detail and a long-window path captures broader temporal context, outperforming single-path CNN, LSTM, Hybrid, and ResNet baselines on OPPORTUNITY and UniMiB-SHAR (Long et al., 2019). These models established a recurring Neural-HAR theme: temporal scale separation and fusion location can matter as much as raw depth.

Hybrid CNN–recurrent systems remained influential, but later work showed that conventional assumptions imported from other sequence domains do not always hold. “Improving Deep Learning for HAR with shallow LSTMs” reports that replacing DeepConvLSTM’s 2-layer LSTM with a 1-layer LSTM increases performance by up to $11.7\%$ for F1-score, reduces learnable LSTM parameters by about $62\%$–$63\%$, and decreases training time by as much as $48\%$ (Bock et al., 2021). This directly challenges the idea that stacked recurrent depth is intrinsically necessary for sensor HAR.

Transformer adaptations introduced a different architectural synthesis. HART is a lightweight sensor-wise Transformer in which accelerometer and gyroscope streams are patch-embedded separately and attention is applied sensor-wise, reducing attention complexity from $O(N^2 d)$ to $O(2(N^2 d/4))$ while improving over ViT-like baselines and showing better robustness to heterogeneous devices and on-body positions (Ek et al., 2022). In multimodal federated HAR, MAGNET treats modality embeddings as graph nodes, constructs a hybrid dynamic-plus-learnable adjacency matrix, and applies graph attention with sparse top-$k$ mixture-of-experts routing before final classification (Bandyopadhyay et al., 3 Aug 2025). In structured industrial settings, a different decomposition appears: a temporal CNN predicts semantic movement attributes, and shallow classifiers such as QDA, HMM, and RF map those attributes—with optional process-step context—to activity labels, improving over stricter attribute-decoding schemes (Lüdtke et al., 2021). Across these lines, Neural-HAR repeatedly favors modular decompositions that reflect sensor structure, temporal scale, or context priors rather than treating activity windows as generic sequences.

3. Learning with scarce labels, imbalance, and weak supervision

A persistent obstacle in Neural-HAR is that realistic activity data are label-scarce and class-skewed. A*HAR formalizes this as a benchmark for semi-supervised learning on wearable wrist data with 24 activities, a maximum-to-minimum class-cardinality ratio of 28, and balanced labeled subsets combined with imbalanced unlabeled pools (Narasimman et al., 2021). The benchmark uses a Mean Teacher framework with a 13-layer convolutional neural network, subject-wise temporal splitting, and labeled budgets of 288, 600, and 1,200 samples. Mean Teacher improves accuracy from $39.13 \pm 2.14\%$ to $41.79 \pm 2.76\%$ at 288 labels, from $11.7\%$0 to $11.7\%$1 at 600 labels, and from $11.7\%$2 to $11.7\%$3 at 1,200 labels, but degrades consistently under class imbalance, which the authors attribute to the method being “naturally not aware of the unbalanced class distribution” in the unlabeled pool (Narasimman et al., 2021). A notable limitation is that the benchmark reports accuracy only, not macro-F1 or balanced accuracy.

Self-supervision addresses the same bottleneck from a different angle. On month-scale Project Baseline Health Study smartwatch data, a convolutional encoder is pretrained on about 42,000 hours of unlabeled wrist accelerometer recordings using a coincidence-learning objective over temporally proximate and augmented windows, then frozen and evaluated with logistic regression on HMPADL, PAMAP2, MHealth, and Daily Sports (Sridhar et al., 2021). The learned 256-dimensional embedding improves over an 8-feature statistical baseline from 70.9 to 79.5 on HMPADL, from 74.3 to 83.3 on PAMAP2, from 82.4 to 93.4 on MHealth, and from 72.8 to 91.1 on Daily Sports (Sridhar et al., 2021). The same work adds an unsupervised salient-activity segmentation procedure in embedding space, which achieves event-level recall/precision of $11.7\%$4 and window-level recall/precision of $11.7\%$5 on continuous PAMAP2 streams, substantially improving low-label HAR on continuous data (Sridhar et al., 2021). This suggests that Neural-HAR performance in free-living settings depends not only on representation learning but also on segmentation quality.

Context can also function as a weak supervisory signal. In logistics-process HAR, process-step labels from a business process model change class priors and transition dynamics; incorporating them through shallow classifiers on top of neural attribute embeddings improves performance even when the process-step information is only partially correct (Lüdtke et al., 2021). A plausible implication is that low-label Neural-HAR need not rely exclusively on more powerful backbones; context-aware decision layers and segmentation can be equally consequential.

4. Efficiency, compression, and edge deployment

One of the defining trajectories of Neural-HAR is the shift from benchmark accuracy toward hardware-constrained deployment. On microcontrollers, quantized 1D CNNs with adaptive inference produce a large Pareto set of models spanning more than one order of magnitude in memory, latency, and energy; reported memory occupation varies in $11.7\%$6–$11.7\%$7 kB, energy consumption in $11.7\%$8 and $11.7\%$9 $62\%$0J, and all models remain compatible with real-time on-device HAR with inference latency $62\%$1 ms (Daghero et al., 2022). The same study shows that sub-byte and mixed-precision quantization can reduce memory by up to $62\%$2 relative to 8-bit models with small or no accuracy loss, but that sub-8-bit arithmetic is mainly a memory optimization unless the target MCU supports it efficiently (Daghero et al., 2022).

Binary neural networks push this logic further on low-power RISC-V processors. Ultra-compact 1D BNNs for HAR replace arithmetic-heavy inference with XNOR-popcount kernels and achieve the same accuracy of a Random Forest with either less memory, up to $62\%$3, or more energy-efficiency, up to $62\%$4, depending on the task and feature-complexity regime (Daghero et al., 2022). The systems contribution is as important as the model family: the paper shows that HAR often occupies an “ultra-compact” regime with only 2–8 channels per layer, making conventional multiple-of-32 BNN libraries wasteful (Daghero et al., 2022).

At the extreme deployment limit, the ISPU study “In-sensor 24 classes HAR under 850 Bytes” demonstrates a 24-class HAR pipeline on an Intelligent Sensor Processing Unit with $62\%$5 B stack usage, but its flagship deployed model is a feature-based XGBoost rather than an end-to-end neural network (Benmessaoud et al., 13 Feb 2025). The contrast is instructive: a 1D ResNet reaches $62\%$6 on raw data and a small MLP on engineered features reaches $62\%$7, yet the sub-850 B deployment uses reduced-feature XGBoost at $62\%$8 accuracy and $62\%$9 KB stack (Benmessaoud et al., 13 Feb 2025). This clarifies the current gap between mainstream neural HAR accuracy and true in-sensor constraints.

A different route to extreme efficiency is to change the computational substrate entirely. Differentiable Weightless Neural Networks use LUT-based inference rather than MAC-heavy layers and reach $63\%$0 and $63\%$1 on UCI-HAR while consuming only $63\%$2 nJ and $63\%$3 nJ per sample, with $63\%$4 ns per sample on FPGA; the reported comparison claims up to $63\%$5 energy savings and $63\%$6 memory reduction relative to prior deep HAR methods (Bacellar et al., 13 Feb 2025). In radar HAR, the system specifically titled “Neural-HAR” reaches $63\%$7 accuracy with 2,719 parameters and $63\%$8 M FLOPs on UoG2020, then maps to a Xilinx Zynq-7000 Z-7007S implementation with $63\%$9s latency and 15 mW dynamic power using zero DSPs and zero BRAM (Wu et al., 26 Oct 2025). Together these results show that Neural-HAR has become inseparable from hardware co-design.

5. Spiking and neuromorphic Neural-HAR

Spiking neural networks form a distinct subfield within Neural-HAR. A wearable HAR study based on SpikeCNN and SpikeDeepConvLSTM replaces conventional nonlinearities with Leaky Integrate-and-Fire neurons whose membrane dynamics follow

$48\%$0

with thresholded spike generation and hard- or soft-reset updates; the models are reported to be on par with ANNs in accuracy while reducing estimated energy consumption by up to $48\%$1 across UCI-HAR, UniMB SHAR, and HHAR (Li et al., 2022). The same study shows that the decay factor $48\%$2 is the dominant spiking hyperparameter: $48\%$3 consistently underperforms $48\%$4, while overly persistent settings can collapse performance on some datasets (Li et al., 2022).

A second spiking direction reinterprets structured recurrent memory through neuromorphic dynamics. The L$48\%$5MU paper converts the LMU’s Legendre-memory mechanism into populations of LIF or CuBa-LIF neurons and adds a learnable spiking front-end so raw smartwatch IMU signals can be converted internally into spikes without a separate handcrafted encoder (Fra et al., 2024). On a 7-class smartwatch subset of WISDM, the Leaky L$48\%$6MU reaches median test accuracy $48\%$7, and the compressed version reaches $48\%$8, while running in under 300 ms on STM32MP1 and Raspberry Pi boards (Fra et al., 2024). This suggests that native spiking recurrent HAR is feasible even on ordinary Linux ARM edge devices.

Spiking Neural-HAR has also been explored in RF sensing. The CSI-based thesis “Approaches to human activity recognition via passive radar” uses a temporally structured LIF network trained with surrogate gradients, reporting average accuracy $48\%$9 for the SNN and $O(N^2 d)$0 for the CNN on the seven-class CSI benchmark; a neurosymbolic SNN variant reaches $O(N^2 d)$1, while DeepProbHAR remains stronger at $O(N^2 d)$2 (Bresciani et al., 2024). The same work explicitly reports poor generalization across subjects, rooms, and days, underscoring that neuromorphic computation does not by itself resolve distribution shift (Bresciani et al., 2024).

6. Robustness, misconceptions, and open problems

Robustness has emerged as a central corrective to benchmark-driven Neural-HAR. A systematic study of wearable IMU variability shows that subject, device, position, and orientation shifts affect deep HAR models differently, that device variability produces the largest performance drops, and that Maximum Mean Discrepancy is generally inversely related to model performance across HARVAR and REALDISP (Khaked et al., 14 Mar 2025). Orientation variability has the smallest impact, while device mismatch—especially with sampling-rate differences—can cause severe degradation (Khaked et al., 14 Mar 2025). This directly refutes the common but weakly supported assumption that high lab accuracy implies deployment readiness.

A second misconception concerns model depth. DeepConvLSTM-style systems are often treated as if stacked recurrent layers were intrinsically beneficial, yet the shallow-LSTM study reports the opposite for five public HAR datasets (Bock et al., 2021). A third concerns terminology around LLMs: the federated multimodal system MHARFedLLM uses a customized T5 encoder-only architecture for time-series tokens inside a larger multimodal graph-fusion framework rather than a general-purpose LLM doing activity reasoning from text prompts (Bandyopadhyay et al., 3 Aug 2025). A fourth concerns archival reliability. One supplied arXiv record, “Spatiotemporal Radar Gesture Recognition with Hybrid Spiking Neural Networks: Balancing Accuracy and Efficiency” (Mazzieri et al., 27 Sep 2025), is described in the provided material as not being the claimed research paper at all but an IOP LaTeX author-guidelines template; consequently, specific technical claims associated with that record are not extractable from the supplied content (Mazzieri et al., 27 Sep 2025).

Open problems recur across the literature. Semi-supervised HAR still lacks class-imbalance-aware methods that do not assume prior knowledge of the unlabeled class distribution (Narasimman et al., 2021). Multimodal federated HAR still leaves missing-modality handling, communication efficiency, and stronger personalization largely open (Bandyopadhyay et al., 3 Aug 2025). RF and CSI systems continue to show poor generalization under subject, room, and day shifts (Bresciani et al., 2024). Radar edge systems remain compact and fast, but lower-precision quantization, multi-radar fusion, event-driven streaming, and on-chip learning are still identified as future directions (Wu et al., 26 Oct 2025). Taken together, these works suggest that Neural-HAR is no longer defined primarily by replacing handcrafted features with deeper networks; it is defined by the joint problem of representation learning, distribution shift, sensor heterogeneity, and deployability under strict resource budgets.