Microwatt power hardware implementation of machine learning algorithms on MSP430 microcontrollers
- Resource Type
- Conference
- Authors
- Heller, Simon; Woias, Peter
- Source
- 2019 26th IEEE International Conference on Electronics, Circuits and Systems (ICECS) Electronics, Circuits and Systems (ICECS), 2019 26th IEEE International Conference on. :25-28 Nov, 2019
- Subject
- Bioengineering
Components, Circuits, Devices and Systems
Computing and Processing
General Topics for Engineers
Geoscience
Robotics and Control Systems
Signal Processing and Analysis
Machine learning
Ultra-low power
Decision trees
Convolutional neural networks
Microcontroller
Seizure detection
- Language
Defining characteristics of Machine learning are its ability to classify big data sets and to learn complex tasks as well as its required high computational power. In this paper we present a hardware implementation of widely used machine learning algorithms on an ultra-low power microcontroller (MSP430, Texas Instruments) to detect epileptic seizures in recorded brain signals. Later, these detection algorithms are integrated in an implantable closed-loop neurostimulator. By outsourcing mathematical operations from the microcontroller's CPU to its integrated hardware modules a power consumption of $82\ \mu \mathrm{W}$ for a decision tree ensemble and $802\ \mu \mathrm{W}$ for a convolutional neural network is achieved.