Oscillating neural networks for low power analog edge AI
A European project including IBM and BMW is developing a new type of analog neural network with phase change materials to reduce the power consumption of machine learning at the edge of the network.
The new technique couples oscillating neural networks (ONN) with phase change materials and could reduce power consumption by a factor of 100 to 1000 say the researchers in the PHASTRAC (Phase Change Materials for Energy Efficient Edge Computing) project.
- In-memory neuromorphic AI slashes power consumption
- imec, Leti form European hub for AI
- ST hints at analog in-memory computing chip
The project, which started last month, aims to develop an analog neuromorphic computing approach based on oscillatory neural networks (ONNs) that seamlessly interfaces with sensors and processes analog data without any analog-to-digital conversion.
The oscillating neurons will be implemented with vanadium dioxide (VO2) phase change material coupled with synapses implemented with bilayer resistive RRAM memories using molybdenum and hafnium dioxide (Mo/HfO2).
The project aims to develop new devices for implementing the ONN architecture and processing the analog sensor data. It is led by researchers from Technical University in Eindhoven, who held a session on ONN technology at the European Conference on High-performance Embedded Architecture and Compilation (HiPEAC) Conference last month in Toulouse. They are working with researchers from BMW, IBM Research in Zurich and the Pázmány Péter Catholic University in Budapest.
Related analog AI articles
- ARM, IBM team on low power analog AI chip
- Analog-in-memory AI processor startup uses memristors
- Neuromorphic analog processor test chip embodies tinyML
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
