Memristors (or memory-resistor) are one of the most promising candidates for next-generation memory applications owing to their high speed switching, ultra low power (non-volatile) operation, ease of fabrication and simple two-terminal structure that enables very high density integration.
In recent years, memristor performance has advanced considerably. Very high levels of endurance (120 billion cycles) and retention have been achieved (15 years), and ultra-high density cross bar arrays have been realized with scalability down to ~2 nm.
Memristors also directly emulate the chemical and electrical switching properties of biological synapses, i.e. the key learning and memory components of the human brain. Memristors exhibit habituation phenomenon, including potentiation, depression and other forms of neural learning such as spike timing dependent plasticity (STDP). Simple neural circuits have been made that can perform logical reasoning, pattern recognition and robotic control.
This combination of properties in a single device (i.e. synaptic functionality, fast switching, low power operation and ultra high-density integration) is very unique and perfectly suited to neuromorphic computing. It is now reasonable to imagine a whole new era of low power yet highly intelligent systems for A.I. applications ranging from smarter phones with cognitive abilities and natural language processing to autonomous vehicles and smarter medical devices.
Our research focuses on memristors that can as well switch by optical means. The use of light as an additional degree of freedom with which devices can be controlled has many advantages. In addition to the much faster speed and higher bandwidth that light offers, the spatial and temporal control of memristor neural networks by light has the potential to give memristor based neuormorphic systems a much greater level of dynamic complexity that increases their ability to learn both better and faster. It also provides task specific learning through light-tuneable synaptic plasticity. By studying new neuromorphic architectures we also hope the research will give new insight into how biological brains function and evolve from birth to maturity.
Comparison of Synaptic Efficacy in Hippocampal neurons (left plot, from Q. Bi and M. M. Poo, J neurosci, vol. 18, pp. 10464, 1998) and our Artificial Synaptic Devices (right). The curves on the right importantly show the ability for dynamic tuneable learning modulated by optical means.