Source-linked AI summary
Adaptive Extreme Edge Computing for Wearable Devices
Erika Covi, Elisa Donati, Hadi Heidari, David Kappel, Xiangpeng Liang, Melika Payvand, Wei Wang
TL;DR
Wearable edge systems need adaptive learning under tight power, computing, communication, and hardware constraints. The paper reviews sensors, biologically plausible continual-learning algorithms, neuromorphic CMOS processors, and memristive devices as a co-designed pathway. It concludes that local adaptive learning, encoding, and device integration remain open requirements for competitive wearable intelligence.
Problem
Wearable systems must process growing sensor data with limited power, computing capacity, transmission efficiency, and latency, while supporting continual adaptation to users.
Method
The paper provides a holistic review spanning wearable sensors, biologically plausible continual-learning algorithms, CMOS neuromorphic processors, and emerging memristive devices.
Results
The review identifies a co-designed platform combining sensors, front-end circuits, neuromorphic processors, and memristive devices as the route toward wearable edge intelligence, while reporting neuromorphic biomedical-signal demonstrations including 91% and 95% anomaly-detection accuracy.
Takeaways & Limitations
Future wearable edge systems require improved local adaptive-learning rules, streamless sensor-to-chip encoding, and efficient event-based algorithms that can approach Backprop accuracy without significantly increasing energy cost.
Takeaways & Limitations
Backpropagation remains difficult for neuromorphic hardware because locking impedes online distributed updates and spiking outputs are non-differentiable, while continual learning with memristive devices remains largely unexplored.
Abstract
from arXiv · showhide
Wearable devices are a fast-growing technology with impact on personal healthcare for both society and economy. Due to the widespread of sensors in pervasive and distributed networks, power consumption, processing speed, and system adaptation are vital in future smart wearable devices. The visioning and forecasting of how to bring computation to the edge in smart sensors have already begun, with an aspiration to provide adaptive extreme edge computing. Here, we provide a holistic view of hardware and theoretical solutions towards smart wearable devices that can provide guidance to research in this pervasive computing era. We propose various solutions for biologically plausible models for continual learning in neuromorphic computing technologies for wearable sensors. To envision this concept, we provide a systematic outline in which prospective low power and low latency scenarios of wearable sensors in neuromorphic platforms are expected. We successively describe vital potential landscapes of neuromorphic processors exploiting complementary metal-oxide semiconductors (CMOS) and emerging memory technologies (e.g. memristive devices). Furthermore, we evaluate the requirements for edge computing within wearable devices in terms of footprint, power consumption, latency, and data size. We additionally investigate the challenges beyond neuromorphic computing hardware, algorithms and devices that could impede enhancement of adaptive edge computing in smart wearable devices.
1 Introduction
Wearable sensing is constrained by computing, power, data-transfer, and latency demands, motivating adaptive extreme-edge processing. The review surveys biologically plausible learning, neuromorphic CMOS systems, and emerging memristive devices as components of a holistic solution.
- Motivation: Wearable devices monitor physiological and behavioral signals, but limited computing capability, high power consumption, large data volumes, and slow transmission constrain conventional systems.Conventional solutions often transmit collected data to external servers for off-chip processing, creating an information bottleneck.
- Neuromorphic principles: Event-based processing, in-memory computing, and massive parallelism form the three principles underlying neuromorphic computing for real-time sensory processing.Spiking neural networks support asynchronous communication, while in-memory computing reduces memory–processor data movement.
- Hardware bottlenecks: Fully CMOS neuromorphic platforms face volatile, area-intensive SRAM and impractical long time constants for always-on adaptive learning.Large networks may require tens of minutes to restore configuration at startup, while long time constants require large-area capacitors.
- Emerging devices: Emerging memristive devices offer non-volatile retention above 10 years, reconfigurability at ≤3.3 V, and volatile time constants from tens of milliseconds to seconds.These properties are paired with small footprint and power efficiency, but require biologically plausible continual-learning algorithms.
- Review scope: The review organizes the design challenge across wearable sensors, continual-learning algorithms, CMOS neuromorphic processors, and emerging memristive devices.It argues that co-designing these categories is key to future smart sensing systems.
2 Wearable sensors
Wearable systems combine diverse sensors and signal-processing stages to capture and interpret human-body information. Their practical deployment is limited by power, computing, communication, privacy, and latency constraints, motivating local edge intelligence.
- Sensor systems: Wearable sensors capture physiological, biochemical, movement, and mental-state information, but require further processing to make signals meaningful for personalized devices.Readout circuits amplify and condition weak, noisy signals before higher-level processing and actuation.
- Wearable modalities: ECG, EEG, EOG, EMG, PPG, and BIS provide complementary measurement modalities spanning cardiac, neural, ocular, muscular, optical, and tissue-impedance signals.Their signal characteristics and sensing mechanisms differ substantially across applications.
- Multisensory integration: Multisensory systems combine sensor types to provide more dimensions of input when an individual sensor cannot satisfy accuracy or robustness requirements.Different sensors return different data that can jointly map an object onto a data stream.
- Edge-computing challenges: Cloud processing introduces wireless-module complexity, power consumption, data volume, transmission-range limits, privacy risks, and communication latency.These drawbacks strongly limit wearable-sensor applications.
- Local intelligence: Neuromorphic hardware is presented as a route toward more power-efficient local processing than von Neumann implementations.The paper connects this direction to brain-inspired processing and wearable edge computing.
3 Models for biologically plausible continual learning
The paper frames neuromorphic edge learning as a way to bring modern machine-learning capabilities to constrained devices. It emphasizes methods that can operate efficiently on neuromorphic hardware.
- Section overview: The section introduces methods for porting modern machine learning to neuromorphic edge devices.It motivates this direction by noting the progress of machine learning and the brain-inspired organization of artificial neural networks.
3.1 Brain-inspired learning algorithms for neuromorphic hardware
Backpropagation is effective for training neural networks but is difficult to deploy efficiently on neuromorphic hardware because of locking and non-differentiable spikes. The section reviews biologically inspired approximations and recurrent-learning challenges for edge systems.
- Backpropagation: Backpropagation trains networks through forward propagation, loss computation, backward error propagation, and parameter updates.It is an iterative, gradient-based, supervised learning algorithm.
- Biologically inspired models: Figure 2 presents e-prop, synaptic sampling, and selective slowing of weight changes as biologically inspired approaches to recurrent learning, sparse robust configurations, and reduced forgetting.The figure describes e-prop as approximating backpropagation through time with random feedback and synaptic sampling as exploiting learning-rule variability and task redundancy.
- Neuromorphic limitations: Locking prevents efficient online distributed implementation because weights await complete forward propagation, loss evaluation, and backward gradient propagation before updating.Backpropagation is also poorly suited to spiking neural networks because their output functions are non-differentiable.
- Gradient approximations: Random feedback alignment replaces non-local error backpropagation with a fixed random feedback network that supplies local feedback to synapses.The feedback weights are excluded from training.
- Spiking sequence learning: Superspike uses error-modulated local learning and a surrogate gradient to handle non-differentiable spike events while learning complex sequences.Its auxiliary function replaces infinitely steep spike events at their timing points.
- Recurrent learning: Backpropagation through time unfolds recurrent networks and stores the extended temporal structure, creating severe power and area challenges for edge computing.Recurrent learning is important because edge-computing tasks often involve temporal dimensions.
3.2 Efficient learning under stringent memory constraints
Neuromorphic edge systems must use limited memory efficiently, motivating sparse connectivity methods that reduce synaptic-weight storage while preserving learning performance. Approaches range from post-training pruning to online rewiring and dynamic regrowth.
- Neuromorphic systems limit resources to improve energy efficiency, making synaptic weights a primary target for memory-footprint reduction.Many network connections converge near zero, motivating sparsity-based methods.
- Post-training pruning commonly removes the weakest weights by magnitude, with some methods reporting over 95% sparsity and negligible performance loss.
- Online stochastic rewiring maintains a fixed number of synaptic connections during learning and reaches close-to-state-of-the-art performance at up to 98% sparsity.
- Sparse evolutionary training prunes small weights and regrows connections at random locations, while related methods use prune-redistribute-regrowth cycles or saliency-based pruning.
3.3 Open challenges and future work
Open challenges for adaptive edge learning include energy-efficient continual operation, robustness to device variability, catastrophic forgetting, and transfer across sensors and tasks. Existing algorithmic and hardware approaches address parts of these challenges but leave future research directions.
- The e-prop algorithm implemented on SpiNNaker 2 achieved an additional two-orders-of-magnitude energy reduction compared with an X86 implementation.
- Edge learning must process continuous data streams under stringent memory and power constraints while remaining robust to variability and noise in devices such as memristors.
- Biologically inspired models treat neuronal and synaptic variability as potential computational resources for stochastic sampling and learning complex statistical dependencies.
- Always-on learning with limited memory can cause catastrophic forgetting, where training additional patterns interferes with previously stored information.
- Transfer learning reuses knowledge from related tasks so new skills can be learned without interfering with previously acquired skills.
- Learning-to-learn models allow networks to modify their own weights and potentially implement optimization algorithms through an outer-loop learning network.
4 Signal processing for wearable devices on neuromorphic chip
Neuromorphic processors provide event-based, parallel, and locally stored computation for wearable biomedical sensing, with platforms demonstrating low-power signal-processing applications. Their practical deployment still depends on local adaptation, feasible on-chip learning, and suitable weight storage.
- Neuromorphic processors: Neuromorphic engineering targets analog/digital processors that emulate biological neurons and synapses using massively parallel, event-based computation and distributed local memory.
- Neuromorphic processors: The reviewed neuromorphic platforms include TrueNorth, SpiNNaker, Loihi, DYNAP-SE, ODIN, and MorphIC, spanning digital, mixed-signal, asynchronous, and configurable architectures.
- Biomedical signal processing: Biomedical applications include EMG, ECG, and intracranial EEG processing, with DYNAP-SE achieving 84% classification using an SVM readout and 74% using a spiking learning method at about 0.05 mW.
- Biomedical signal processing: For hand-gesture classification, Loihi and ODIN/MorphIC achieved the same accuracy as an embedded GPU with energy-delay products 30x and 600x more efficient, respectively.
- Adaptation: Local adaptation is relevant because biomedical signals vary across people, and neuromorphic chips can support online, on-edge adaptation for personalized medicine.
- Adaptation: Current on-chip learning platforms implement rules including STDP, reinforcement learning, three-factor learning, SDSP, and configurable algorithms on SpiNNaker.
- Open challenges: Online on-chip learning is constrained by the locality of weight updates and by weight storage requirements, including non-volatility, linear updates, and analog states.
5 Memristive devices and computing
Memristive devices are presented as scalable, low-power, CMOS-compatible candidates for extreme-edge neuromorphic computing. The section outlines their potential and identifies integration with CMOS and sensors as a central future challenge.
- Memristive devices attract attention for scalability, low-power operation, CMOS compatibility, and volatile/non-volatile properties.
- Figure 3 surveys RRAM, phase change memory, MRAM, and FTJ device structures used in neuromorphic computing.
- The review examines memristive technology’s current challenges and future perspectives for neuromorphic computing.
5.1 Conventional and wearable memristive devices
Memristive devices store resistance states through diverse physical mechanisms and can support energy-efficient, low-latency processing near wearable sensors. Their volatile and flexible implementations offer routes toward local computation and biological short-term-memory emulation.
- Memristive devices change and memorize resistance states through forms including RRAM, PCM, MRAM, and FTJ.
- Memristive learning can reduce system size, processing latency, energy use, and sensor-data bandwidth by processing information where it is acquired.
- Volatile RRAM behavior lasting tens of milliseconds to seconds can mimic the timescale of biological synaptic short-term memory.
- Memristive devices can be fabricated on flexible substrates, enabling possibilities for local computation within wearable devices.
- Table 2 summarizes key features of non-volatile memristive devices.
5.2 Memristive devices for neuromorphic computing
Memristive devices can implement both artificial neurons and synapses, supporting brain-like and deep-learning neuromorphic architectures. Reported applications include pattern recognition, collective decision making, speech processing, and biomedical signal analysis.
- Memristive neurons can emulate integrate-and-fire behavior and be more compact than CMOS neuron implementations.
- Memristive synapses implement tunable weights through resistance or conductance changes and support learning rules such as STDP.
- Memristive neuromorphic systems span deep-learning accelerators and brain-like computing based on spike representations and collective decision making.
- Figure 4 depicts memristive devices functioning as threshold-based neurons and artificial synapses.
- Coupled memristive oscillator networks have emulated attention-dependent selection of relevant information from multiple sensor inputs.
- Memristive technologies have been investigated for speech, biosignal, medical-image, and cancer-diagnostic applications.
5.3 Open challenges and future work
Open challenges include device non-idealities, incomplete integration of memristive devices with CMOS systems, and the demands of continual learning under wearable-device constraints. Future work must address robust learning and co-integrated end-to-end processing.
- No single memristive device simultaneously satisfies all requirements for mainstream backpropagation, including linear reading, analog tuning, symmetric updates, retention, and endurance.
- Readout and weight-update techniques have been proposed to compensate for nonlinear and asymmetric device behavior.
- Memristive stochasticity and noise can also be exploited for probability computation and escaping local minima in brain-like networks.
- Co-integrated memristive neuromorphic processors remain under development despite existing CMOS integration demonstrations.
- A proposed integration path begins with non-volatile devices and peripheral circuits, including logic and multiply-and-accumulate operations, before combining different technologies.
- Continual learning in wearables must preserve learned features as new experiences arrive, avoiding catastrophic forgetting under limited memory capacity.
- A hybrid convolutional ANN, SNN, and memristive-device system was reported as robust against catastrophic forgetting with 93% accuracy across trained and non-trained classes.
6 Discussion and Conclusions
The paper frames adaptive wearable edge computing as a system-level challenge requiring coordinated sensing, processing, memory, and continual learning. It concludes that co-designing these elements is necessary because no single technology currently provides an ideal solution.
- 6 Discussion and Conclusions: The study surveys core elements for wearable devices with extreme edge adaptive computing, including sensors, learning algorithms, neuromorphic processors, and emerging memory.Its scope spans bio-signal sensing, biologically plausible continual learning, CMOS processing, and memristive devices.
- 6 Discussion and Conclusions: Wearable edge processing must jointly address sensing diversity, limited power, fast response, and classification accuracy rather than relying solely on cloud processing or energy–capacity trade-offs.The paper calls for optimization across learning algorithms, architectures, and other platform elements.
- 6 Discussion and Conclusions: Continual learning is required for adaptive wearables, but robustness to hardware non-idealities, forgetting, task-to-task information transfer, and spike encoding remain unresolved.SNNs support low-power in-memory computing and fast real-time processing, while their input encoding remains challenging.
- 6 Discussion and Conclusions: The envisioned platform combines embedded sensors, front-end interfaces, neuromorphic processors, and memristive devices to move intelligent algorithms to the wearable edge without significantly increasing energy cost.The paper also identifies improved local learning rules and streamless sensor-to-chip encoding as needed developments.