Streaming wearable sensors create continuous, high-volume signals that must be reduced on-device to save energy, protect privacy, and work over limited networks. Research on representation learning and sparse sampling provides the strongest evidence for which architectures perform best in this setting. Geoffrey Hinton University of Toronto demonstrated that deep autoencoders can learn compact, task-relevant representations of raw signals, making them natural candidates for on-device compression. Emmanuel Candès California Institute of Technology and David Donoho Stanford University established the theory of compressive sensing, which leverages signal sparsity to reconstruct signals from fewer measurements and can be implemented with lightweight sampling front ends.
Architectures suited to on-device streaming
In practice, 1D convolutional autoencoders and lightweight temporal convolutional networks offer a strong balance of compression ratio, latency, and energy cost for wearable streams such as accelerometry and photoplethysmography. These convolutional architectures exploit local temporal structure and are easier to quantize and prune for microcontroller deployment than large recurrent or transformer models. For medically structured signals like ECG, variational autoencoders add principled uncertainty estimation that improves downstream clinical safety. Where signals are known to be sparse in a transform domain, implementations based on compressive sensing reduce sampling and transmission overhead by design. For on-device efficiency, the hardware-aware guidance of Vivienne Sze Massachusetts Institute of Technology emphasizes co-design of model and accelerator to minimize memory transfers and energy per inference. Pete Warden Google advocates TinyML practices that combine small architectures with post-training quantization and entropy coding to produce compact bitstreams for transmission.
Relevance, trade-offs, and consequences
Choosing an architecture depends on sampling rate, signal modality, and the acceptable reconstruction error for the application. Higher compression can save battery and network cost but may remove clinically or behaviorally relevant nuance. Consequences of poor choices include degraded recognition accuracy, biased outputs across demographic groups, and regulatory risk for health applications. Cultural and territorial realities matter because connectivity and device lifecycles differ globally; on-device compression that reduces reliance on cloud infrastructure can improve access in low-bandwidth regions and lower the environmental footprint of continuous monitoring. For most wearable streaming tasks, a compact 1D convolutional autoencoder or TCN combined with pruning, quantization, and entropy coding, implemented with TinyML and hardware-aware practices, offers the best practical trade-off between compression, latency, and reliability.