Professors Joshua Yang and Qiangfei Xia of the Electrical and Computer Engineering Department at UMass Amherst led a research team from multiple institutions – including the NASA Ames Research Center, Hewlett-Packard Laboratories, and the Air Force Research Lab – which has realized the first “capacitive neural network” experimentally, a leap forward in the development of a new neuro-biological architecture that can mimic very useful qualities of the human brain and nervous system.
The results were published on August 10 in Nature Communications under the title of “Capacitive neural network with neuro-transistors.” The research described in Nature Communications demonstrated an alternative solution for neuromorphic computing; that is, the use of capacitive coupling (vs. resistive coupling) integration systems containing electronic analog circuits to mimic neuro-biological architectures present in the human brain and nervous system.
Yang and Xia stated that electrical and computer engineers have long been trying to duplicate the computing archetype of the brain, which, unlike CMOS circuits, is not limited by the separation of memory and processing, serial execution, power inefficiency, and programming-intensive issues of the von Neumann architecture.
Several emerging devices are a potential route to neuromorphic computing that can mimic functionalities of neurons and synapses more efficiently than traditional CMOS circuits and thus provide more capable and efficient neuromorphic systems.
In their Nature Communications paper, the authors explained that one of these promising emerging devices is the memristor (often called RRAM when used for memory), a resistor whose resistance depends on the history of the applied electrical bias.
“Being extensively studied in the last decade,” said the authors, “the memristor has been identified as a leading candidate for neuromorphic computing because it can be used to build artificial synapses and artificial neurons. Such properties lead to the demonstration of full-memristor neural networks with learning capabilities based on resistive coupling among the devices.”
However, as in all resistive networks, the resistive devices are passive elements, which dissipate the energy of signal propagating in the networks in the form of heat without recycling them. These interrelated factors may compromise the energy efficiency of resistive neural networks.
A new type of capacitors that are electrically programmable and possess memory are called memcapacitors. As the authors stated, “A capacitive neural network enabled by memcapacitive elements has functions based on capacitive coupling and can potentially be more energy-efficient than traditional memristors since it provides a temporary storage of electrostatic field energy rather than converting it into heat during operations.”
To summarize their research, as described in Nature Communications, Yang and Xia explained that “In this new work, we have for the first time realized a capacitive neural network for neuromorphic computing. In contrast to the prevailing resistive neural network, a capacitive neural network enabled by our newly developed device – a memcapacitor (instead of a memristor) – does not use static current in the circuits and can recycle the electrical energy, which is envisioned to be much more energy-efficient than a memristor-based neural network for computing.”
The authors concluded that “The memcapacitive implementation of a neural network provides an unexplored alternative, as competitive as the resistive neural networks if not more, for hardware implementation of neuromorphic computing. The promise for a potentially better emulation fidelity and improved power efficiency may spur new research directions for memory device and machine learning communities.” (August 2018)