AI That Doesn’t Drain Wearable Batteries

Electronic components such as sensors and microcontrollers have been shrunk down in size and cost to the point where they can practically be incorporated into all sorts of wearable devices. These wearables offer tremendous potential in areas like health monitoring, where they can continuously collect and process data. The insights provided by this information could help health care professionals to diagnose medical conditions earlier, and create more effective treatment plans.

But while data collection with wearable electronics is essentially a solved problem, processing the data still presents many challenges. The nature of health-related data makes it very complex, to the point that developing traditional, hardcoded algorithms is impossible. As such, machine learning algorithms are commonly deployed for these purposes due to their ability to predict and classify complex phenomena.

However, when it comes to the tiny, low-power microcontrollers found in a typical wearable device, these algorithms can quickly overwhelm their modest resources. But now, a new approach developed by researchers at ETH Zurich may help these little processors chew through complex algorithms with cycles to spare. Called NanoHydra, their system is a lightweight and energy-efficient way to run Time Series Classifications (TSCs) on the tiniest of computing platforms.

TSC involves predicting class labels from sequences of time-dependent data, such as electrocardiogram (ECG) signals, brainwave patterns, or accelerometer readings. Conventional deep learning techniques like convolutional or recurrent neural networks can handle such tasks well, but they demand far more memory, energy, and processing power than microcontrollers can provide. NanoHydra overcomes these problems by trimming down the computational complexity of these algorithms without sacrificing accuracy.

The system builds on earlier methods known as ROCKET and HYDRA, which use random convolutional kernels to extract meaningful features from sensor data. NanoHydra streamlines this approach by using binary kernels (simple patterns made up of +1 and −1 values) to replace the floating-point operations that typically bog down small processors. It further substitutes costly mathematical functions, such as square roots and divisions, with lightweight arithmetic shifts that achieve similar results at a fraction of the energy cost.

The researchers implemented NanoHydra on GreenWaves Technologies’ GAP9 microcontroller, an ultra-low-power chip with an eight-core cluster optimized for parallel processing. By spreading out the workload across multiple cores and using SIMD (Single Instruction Multiple Data) operations to process several data points at once, the system performs quite well. It can classify a one-second-long ECG signal in just 0.33 milliseconds while consuming just 7.69 microjoules of energy per inference, making NanoHydra about 18 times more efficient than previous state-of-the-art methods.

Despite its frugal use of resources, NanoHydra doesn’t compromise on accuracy. On the widely used ECG5000 dataset, it achieved 94.47% classification accuracy, rivaling heavyweight desktop-class algorithms. The team estimates that a battery-powered wearable device using NanoHydra could operate continuously for more than four years without recharging. Between the long battery life and accuracy, devices powered by NanoHydra could prove to be very popular with their users.