The human brain contains approximately 100 billion neurons. Inspired from its workings, a team of IIT Delhi researchers led by Professor Manan Suri, Department of Electrical Engineering, have invented a new neuron model named as DEXAT (Double Exponential Adaptive Threshold neuron).
This model when inserted in neuromorphic AI networks for different applications such as data processing and automation, leads far more efficient and accurate results. "Some examples are smart cities, robotics, Internet of things, industry 4.0, security and defence, and healthcare etc.," shares Suri.
As the study has attained positive results not only in the hardware part, but also the theoretical aspect of the technology, it was recently published by the international journal, Nature Communications.
Suri and research students Ahmed Shaban and Sai Sukruth Bezugam had initiated this study towards the end of 2018. "We routinely investigate semiconductor memory technology and neuromorphic hardware, so this topic was a fit for our research," says Suri, who decade-old hand in the field.
Simplifying the heavy jargon for our benefit, Suri explains that the terminology we use is very different from what people use in layman’s language. "For a normal person, a device can be a television, laptop or a smartphone, but for us devices are small tiny entities that are embedded deeply inside a semiconductor chip (not visible to the naked eye). These help the algorithms run for any specific application or system. And by system, it means devices being used by a user. For example, a computer," he says.
The domain of AI is very vast. "One cutting-edge sub-topic of AI that takes inspiration from nature is called neuromorphic computing. As the name suggests, it tries to mimic neural networks inspired from mammal brains. This particular work of ours lies in the sub-field of neuromorphic computing with a focus on hardware," he says.
Energy efficiency is of prime importance. Giving an example, he says, “If a person is using a portable device like a smartphone, our focus is on how fast and efficient the hardware is, no matter what application is run on it.”
Most AI and neuromorphic systems involve a neural network, for tasks such as speech recognition, image processing, Google translate, Siri, or face recognition, for example. "On the theoretical side, we innovated a new neuron model. When we use this neuron model inside recent neural networks, the performance drastically improves on multiple fronts," he adds.
"Coming to the second part of the innovation - hardware, we have demonstrated some specialised nanoelectronic semiconductor devices called resistive memory," he shares. The proposed nanodevice neuromorphic network was found to achieve 94 per cent accuracy.
"We can exploit new theoretical models only by building systems based on advanced nanoelectronic semiconductor devices. We have also filed a patent for the invention. The next step would be to develop full systems (a device in layman's terms) utilising the invention that can interact with live sensors and other peripherals," he concludes.
IN A NUTSHELL
This model when inserted in neuromorphic AI networks for different applications such as data processing and automation, leads far more efficient and accurate results.