Background
Natural Language Processing (NLP), time series analysis and speech recognition are among applications that require specific machine learning techniques for processing temporal data. Reservoir computing (RC) is a machine learning technique that is particularly well-suited for processing temporal data. Typically, the realization of RC systems requires a recurring neural network comprised of layers of volatile and non-volatile devices. Having different devices in different layers adds to the cost, latency, training complexity and footprint of the Reservoir/ Neuromorphic Computing chip.
Description of the invention
Inspired by natural neurons where the transport of Na and K ions forms the basic neural impulses, researchers at the University of Waterloo invented a memristor device that exhibits synapse-like short-term plasticity (STP) behavior. The device is comprised of TiOx, which is known as a good cathode material, and Lithium Phosphorous Oxynitrate (LiPON), a good solid-state Li electrolyte, together forming a device capable of controllable ionic motion under applied voltages, and configurable to either volatile or non-volatile behaviour. The invented universal memristor devices can be used in manufacture of RC/Neuromorphic Computing systems at scale and on the same CMOS layer.
Advantages
- Ease of programing, i.e. no forming process or compliance current is needed for volatile memristors
- High uniformity and great endurance (due to lack of any conductive filaments)
- Volatile analog devices can be configured into non-volatile memory units with multibit storage capabilities
- Single platform of both volatile and non-volatile units for performing neuromorphic training (volatile) and weight storage layer (non-volatile)
Potential applications
- Field programmable neural network arrays (FPNA) for customizable neuromorphic computing tasks
- AI accelerators
- Compute in memory
- Any application that uses recurrent neural networks can benefit from this invention including Natural Language Processing (NLP), time series analysis (stock price prediction, sales forecasting, sensor data analysis), speech recognition, video analysis, medical diagnosis.
- The tuneability of volatile timescales and convertibility from volatile to non-volatile enables novel applications such as fade-able memory for sensitive information distribution

Schematic of the memristor device
Reference
10263
Patent status
Patent pending
Stage of development
Prototype built and tested
Ongoing research
Contact
Scott Inwood
Director of Commercialization
Waterloo Commercialization Office
sinwood@uwaterloo.ca
uwaterloo.ca/research