IIT-B designs electronic neuron to make devices faster

A team of researchers led by the Indian Institute of Technology- Bombay (IIT-B) have designed an electronic neuron that mimics the nerve cells or the neurons of a human brain and can help devices make faster decisions.

The study, published in the journal APL Materials of the American Institute of Physics in September last year, proposes a theoretical framework of neural networks - an active field of research in machine learning - that can enable computers to perform everyday tasks such as image, voice and pattern recognition. In simpler terms, the neurons will improve the results of facial recognition, voice recognition and understanding of languages.

For instance, currently, electronic gadgets such as computers and mobile phones are based on Complementary metal-oxide-semiconductor (CMOS) technology. In this technology, the device will go through millions of possibilities before reaching an answer when prompted for facial recognition. The neural network proposed by the team can bypass the possibilities and reach an optimal answer much more quickly.

“Today, a lot of neural networks are focussed on software that runs on the cloud, which has ample energy to work as they are supported by dedicated server farms,” said Udayan Ganguly, principal investigator and professor from the electrical engineering department, IIT-Bombay. “Unlike most computer hardware designed over the past few decades that served well with simple programs to perform functions such as counting, today complex problems such as searching for an optimal route for a drone need programs with artificial stochastic neurons,” he added.

“Neurons or activation functions are essential components for neural networks, and their realisation such as the one demonstrated in this study, is a valuable development from the perspective of hardware implementation,” said Manan Suri, assistant professor at IIT Delhi and lead, non-volatile memory and neuromorphic research group, who was not a part of the study.

Essentially, the team at IIT-Bombay comprising faculty, undergraduate and postgraduate students

along with engineers from Intel have designed a resistive random access memory (RRAM) that uses artificial neurons to build a network which can help computers solve problems such as finding the optimal route for a drone.

The team considered an existing theoretical framework of neural networks called a Boltzmann machine - a deep learning model to optimise a solution to a problem.

“A Boltzmann machine can enable everyday tasks like image, voice and pattern recognition,” said Ganguly adding that the stochasticity (randomness) in Boltzmann machine results in the ability to statistically estimate the output, which is unnatural for deterministic machines. The project was funded partly by the Department of Science and Technology’s Nano Mission and Ministry of Electronics and IT (MeitY).

The RRAM was built using a crystalline manganite (PrxCa1−xMnO3) which can act as a memory

device. Researchers tested and compared the performance of the newly designed RRAM with that of conventional silicon-based hardware. Results showed their hardware design solved a problem with 98% accuracy and needed just one-tenth of the area of conventional

semiconductor-based hardware. Its power efficiency was also found to be four times better than CMOS.

“This implies that a Boltzmann machine chip, based on RRAM may be computationally more powerful and energy-efficient,” Ganguly said. Researchers have filed for a patent on this work. “The devices are in the experimental stage, but the chip design needs to be implemented in devices. Such systems are of great commercial interest and would be interesting for high-tech start-ups,” he added.