A miniature, super-fast Silicon neuron could make brain-like chips

Brain-like electronic chips are a “holy grail” for computing technology. Recently, a team of Indian researchers has demonstrated an electronic neuron. It is 10x smaller and 1000x faster than biological neurons. A network of such neurons enables biology-like artificial intelligence (AI) in silicon chips. Such AI enables wide range of human-like tasks from ability to learn and then recognize patterns e.g. flowers, faces or even malignant tumors from benign ones.

Researchers around the world are engaged in the scientific pursuit to understand the human brain. Billions of interconnected neurons (the fundamental unit) enable parallel information processing in brain, outperforming today’s super computers. To put it in perspective, super computers take around million Watts of power, whereas a human brain consumes a mere 20 Watts of power to perform similar operations – thus biology a startling million fold more efficient than modern supercomputers!

Today, popular search engine softwares are able to recognize voice and images using traditional digital server farms that guzzle energy. The energy efficiency in biology partly lies in the neurons’ ability to code information in the timing of tiny “voltage spike” rather than digital “1” or ”0” expressed as large and small voltages. Following a unique approach leveraging on the existing fabrication technologies for mass produced high-speed chips, a team of researchers led by Prof. Udayan Ganguly in the Department of Electrical Engineering, IIT Bombay, devised an artificial neuron based on silicon-on-insulator(SOI) transistor technology.

The article published in Scientific Reports on 15th August, 2017 shows that the artificial neuron responds to stimulus akin to biology to produce electrical spikes – with a significant advantage of 10 times size reduction and 1000 times speed-up.

A network of such neurons is able to learn a range of classification tasks. Not only can it learn to classify different variant of Iris flower (Iris Sentosa, Iris Virginica, Iris Versicolor), but also classify malignant / benign cancers. “This study presents an essential stepping stone towards constructing energy efficient, bio-inspired machines that can learn to classify in diverse environments” remarks team member Prof. Nihar Ranjan Mohapatra, IIT Gandhinagar.

Graduate student Sangya Dutta works in the IIT Bombay Nano Fab, where she designs and tests these silicon neurons. Other team members, like graduate student Vinay Kumar design control circuits, and staff, Aditya Shukla, develops classification tests for the neural networks. Sangya explains “Conventional neurons use ions to conduct electricity and produce short “zaps” of current to represent “thoughts” i.e. information. We use electrons instead of 10,000 times heavier Sodium/Potassium ions to enable high-speed.” Inside the artificial neuronal device, a few electrons are driven by electric field to shoot out like bullets to break a tiny fraction of chemical bonds (one Si-Si bond in a billion) temporarily to generate a chain reaction, which delivers the signature neuronal “zap”. “The physics is beautiful!” remarks Prof. Udayan Ganguly.

The diverse team is partially funded by Nano-Mission, Department of Science and Technology to enable the break-through. Another stakeholder is Intel Corporation that has been supporting this breakthrough research under Intel India PhD Fellowship Program.

“These neurons leverage partially depleted Silicon-on-Insulator Technology, platform which is a commercial manufacturing technology,” opines Prof. Suman Datta, who directs a research center on Collective and Neuromorphic Computing at the University of Notre Dame, USA.

It is clear that leveraging a mature platform will support a large neural network – that will take hardware closer to the ultimate benchmark – the human brain with a 100-billion-neuron network.

Image_Neuron

References:

S. Dutta, V. Kumar, A. Shukla , N. R. Mohapatra, and U. Ganguly, “Leaky Integrate and Fire Neuron by Charge-Discharge Dynamics in Floating-Body MOSFET” Scientific Reports 2017, link

U. Ganguly, S. Dutta, V. Kumar “Leaky Integrate & Fire (LIF) Neuron based on Floating Body Effect” (Application No. 201721027169)