Project: AI Hardware Accelerator: Encoding Multi-timescales in Artificial Synapses
Key information:
| Student | Alexander-Hanyu Wang |
| Academic Supervisors | David Thomas, Harold Chong, Ruomeng Huang, Firman Simanjuntak |
| Cohort | 2 |
| Pure Link | Active Project |
The timescales in biological brain not only have inherent temporal properties such as neural and synaptic time constants, but also multiple timescales of learning and adaptation to changes in environment such as timescales of synaptic plasticity. The neural mechanisms underlying the multiple timescales demonstrate the performance of temporal dynamics of brain and behaviour in environment. Therefore, it is crucial to understand time constants and timescales of the neural processing mechanism in biological brain for study of neuromorphic computing in the fields of neuroscience, machine learning and AI electronics.
Memristors can be classed as a new device technology and is known to be the closest electronic component to synaptic function. It has potentiation and depression properties similar to those of biological synapses. These devices are particularly interesting due to their programmability, compatibility as well as their behaviour. Currently, non-volatile characteristics have been studied and exploited such as I-V characteristics and switching properties. However, volatile characteristics are often overlooked and normally regarded as a disadvantage of the device. In order to fully appreciate the potential of volatile devices for use in hardware-based security and mimicking synaptic functions, more research into this field is required.
Volatile memristors exhibit forgetting characteristics which traditional electronic synapses nor traditional non-volatile memristors are capable of demonstrate. Volatile memristor based devices can imitate the forgetting effect of human brain to filter information effectively which allows a stronger learning ability of the neural network.
The aim of this project is to develop and demonstrate hardware-plasticity for AI applications. The development of short-term plasticity through the use different types of volatile memristors can be achieved in the University of Southampton nanofabrication cleanroom. The project will also involve development of novel hardware-plasticity technology through integration of conventional and emerging technology at device level which can further drive the efficiency of AI applications.