Ultra-Low Power Neuromorphic Front End Chip for Wearables
NeuroSense, a Tiny AI chip based on the Neuromorphic Analog Signal Processor, addresses the most annoying problems of mass-market wearables: power consumption, accuracy of heart rate measurements, cloud connectivity and data privacy issues
Mass market wearables could be smarter!
People need them always-on, low-power, accurate and cloud-independent
To support modern Neural Networks’ capabilities, a wearable’s hardware consumes a lot of electricity, because AI requires compute resources and cloud connection. That’s why mass market wearables, such as fitness trackers, smartwatches, health monitoring devices, and remote care wearables for seniors, have short battery life and low tracking accuracy.
1Heart rate calculations, while in motion, have low accuracy due to inconsistent noise factors
2PPG and IMU calculations are resource-hungry
3Human activity learning requires loads of raw data transferred to the cloud
1Three times better heart rate accuracy than algorithm-based calculations
2Ultra-low power consumption below 100µW
3Human activity recognition on sensor level, cloud-independent operation
1 sec: ECG reference, NeuroSense
Heart rate measurement by NeuroSense is twice as accurate as algorithm-based calculations.
NeuroSense offers increased accuracy of heart rate monitoring and the advanced functionality of human activity recognition.
A simple PPG configuration with only two LEDs and one photodiode is enough for the heart rate measurement.
NeuroSense can inform the user about a weak PPG signal based on the heart rate confidence metric value, prompting the user to adjust the wearable device.
Activity matrices allow the wearable device to offer classification-based features:
- Learn and automatically detect new human activities
- Compare the activity execution to a desired reference execution (for example, of the trainee to the trainer’s execution)
- Classify human activity into the base classes like rest or moderate or intensive activity.
NeuroSense outperforms other solutions in the market of affordable wearables.
In typical devices, power savings come at the expense of performance and accuracy limitations. Otherwise, high accuracy may result in low operating times below 20 hours on the battery charge.
In contrast, NeuroSense enables by orders of magnitude longer operation and continuous monitoring and recognition due to analog computations and additional power-saving methods. At the same time, the neural network approach makes biometrics extraction far more accurate than algorithmic methods.
NeuroSense saves the limited space inside a small wearable device.