The human brain is nature's response to the need for information storage and processing. It performs very complex tasks efficiently while consuming very low energy. For this reason, scientists have been trying to improve the performance of traditional computers by reproducing known brain mechanisms. These brain-inspired computers are called neuromorphic systems. To realize such a system, different functional components that make the brain as well as information storage and processing mechanism that happen inside the brain ought to be implemented inside a silicon chip. Interestingly, modern transistors are much smaller and faster than neurons! However, modern computers cannot perform as efficiently as the human brain. The main point is that each neuron make lots of connections with other neurons and form a huge network of computing elements that work together to achieve complex tasks like sensation, perception, and decision making. These connections, called synapses, are very complicated and perform several tasks. In fact, the synaptic strength of synapses can be modified through a learning process called synaptic plasticity which means new information is stored inside the synapse.
Implementation of synapses with electronics is the major obstacle to realizing the neuromorphic chips because transistor-based synapses are very slow, power-hungry, and bulky. Fortunately, with the introduction of the"Memristor", hopes for the efficient realization of neuromorphic systems have been renewed [IEEE1971,Nature2009]. A memristor is a two-terminal electronic device that has non-volatile memory characteristics in the sense that it can retain its resistance for a long time in a non-volatile manner. It has been proved that synaptic plasticity learning mechanism can be realized by the memristors, as a result, they have been widely used as synapses in neuromorphic circuits Nano Lett. 2010.
In neuromorphic systems, unlike traditional computers, memory and CPU are implemented in the same chip, hence, they offer superior performance [ISOCC2020, Neuromorphic Systems]. Thanks to the development of memristors, neuromorphic chips can perform human-like cognitive tasks such as vision, classification, and inference in an inherently energy-efficient way. The Neuromorphic Chip Market is valued at USD 1.952 Billion in 2020, and is expected to reach USD 7.388 Billion in 2026 [NEUROMORPHIC].
Memristors can be realized in transition metal oxides or complex strongly correlated oxides, which exhibit metal-insulator transitions. In this project, we aim to fabricate the memristor with stable analog behavior to emulate neuronal activities, such as leaky integrate fire behavior and synaptic plasticity. Using crossbar technology, an array of memristors will be incorporated into a more complex system to realize a neuromorphic chip Nano Lett. 2012. This research project includes 1) identifying proper material and device structure to achieve a high-quality memristor, 2) design and optimization of the crossbar array with and without selector device to prevent sneak path problem associated with crossbar technology, 3) design analog front-end to read out the information stored in the memristor as well as memristance tuning circuit, 4) design a digital processor inside the same substrate for further signal processing, 5) develop algorithms to perform human-like cognitive tasks utilizing content-based memory and in-memory computing techniques.
Figure 13. Characteristic of memristor
Figure 14. Components: resistor-capacitor-memristor-inductor