NJIT Graduate Student Wins IBM Fellowship to Develop Brain-Inspired Computing Systems
S. R. Nandakumar, a graduate student in electrical engineering, has won a coveted IBM Ph.D. fellowship to support his work on computer systems that mimic the architecture of the human brain. The international fellowship program, which is intensely competitive, awards exceptional graduate students in diverse fields who are tackling technical problems fundamental to innovation.
Nandakumar is developing systems that learn to perform intelligent tasks, from recognizing words and images to executing higher cognitive functions such as speech recognition and language translation. By mimicking the key features of the brain’s network – its neurons, synapses, and mechanisms that change the connections in the network over time – he hopes to raise the level of computers’ learning and inference capabilities while lowering the amount of energy required for their operation.
“The brain is a fascinating information-processing system and we are trying to mimic its behavior to build the next generation of intelligent computing platforms,” Nandakumar says. “We will be implementing new brain-inspired algorithms using IBM's nanoscale hardware technologies.”
Nandakumar joined NJIT in 2016 to work on this project with Bipin Rajendran, (below, right) an associate professor of electrical and computer engineering and expert in nanoscale electronic devices and system design.
“The goal of our research is to build novel computing systems that are inspired by the architecture of the brain,” Rajendran noted. “Compared to the machine-learning algorithms that are popular today, we are developing algorithms for spiking neural networks that mimic the brain’s ability to encode information in the electrical spikes signaled by neurons. We hope that these algorithms, when implemented using nanoscale devices, will one day lead to learning systems that approach the efficiency of the brain.”
The neurons in the brain fire at a very slow rate to communicate with each other compared to the signaling speeds in modern computers. However, each neuron's spike is transmitted to thousands of other neurons. It is believed that this parallel information-processing paradigm allows the brain to operate with very high efficiency: a mere 20 watts is sufficient to control all of a person’s actions, thoughts, and emotions.
Rajendran added, “We are very fortunate that IBM Research, which is the leading research center in the world in the field of cognitive computing, is collaborating with us on this exciting project. IBM's support of Nandakumar will enable us to jointly explore several new promising algorithms on IBM's nanoscale hardware platforms for brain-inspired computing.”
The new algorithms will be implemented on hardware platforms that use nanoscale devices made of chalcogenide materials used for making phase change memories, a novel solid-state memory technology. These devices, unlike the transistors in today's computers, can store information in an analog fashion, significantly improving overall learning efficiency.
"Spiking neural networks are considered to have significant potential for realizing highly energy-efficient neuromorphic hardware. IBM and NJIT are jointly investigating algorithms and non-von Neumann computing platforms to address the challenging problem of learning in spiking networks," said Evangelos Eleftheriou, Ph.D., IBM Fellow, who leads the research group at IBM Research-Zurich.
Nandakumar will spend three months this summer in Zurich working closely with the IBM team, in particular with Abu Sebastian, Ph.D., who also manages a European Research Council project on non-von Neumann computing.
Von Neumann architecture, named for the computing pioneer John von Neumann, separates the two functions of data processing and storage in computing systems today. “But that’s not how the brain works,” Rajendran noted. "IBM’s new hardware platforms, for example, would enable devices to seamlessly merge information processing and storage using substantially less power than today's computers."
But several challenges remain before computers achieve the efficiency of the human brain. "We are currently studying how the systems perform, especially in light of reliability constraints that are inherent to nanoscale devices,” Rajendran says. “The initial results of our research appear promising."
Nandakumar will present the team’s findings at the upcoming 2017 Device Research Conference, to be held in June at the University of Notre Dame in Indiana.