Achieving efficiencies in neural network computing
Neuromorphic computing uses hardware based on the structure, processes and capacities of neurons and synapses found in biological brains. Pioneering scientists believe that this approach could provide an energy-efficient alternative to conventional computing architectures. Potential applications have already been identified in various fields, including artificial intelligence (AI), robotics and autonomous systems. In order to be successfully brought to market at large scale however, certain challenges still need to be addressed.
Overcoming neuromorphic computer limitations
The purpose of the EU-funded PlasmoniAC project was to find ways of overcoming some of these issues. “We wanted to address the inherent efficiency limitations of neuromorphic electronic computational systems,” explains PlasmoniAC project coordinator Nikos Pleros from Aristotle University of Thessaloniki in Greece. “This includes speed and power limitations, as well as difficulties in connecting with electronic analogue processors.” To do this, the project sought to develop materials based on plasmonics. This is a cutting-edge field of research that focuses on the resonant interaction between electromagnetic radiation and free electrons at the interface between a metal and a dielectric material such as air or glass. This interaction generates electron density waves called plasmons, or surface plasmons. The PlasmoniAC project team believed that developing new plasmonic materials and technologies could be the key to optimising the computational power, size and energy of neuromorphic chips, and help to achieve efficiencies in neuromorphic computing.
Experimentally validated plasmonic prototypes
The project team began by investigating the potential of various novel plasmonic-based materials for use in neuromorphic computing. These included materials such as barium titanate, plasmonic organic hybrid and titanium dioxide. Deep learning models were also adapted. The project team developed a new class of optics-informed deep learning models and architectures, that were then validated in PlasmoniAC prototypes. “We were able to assess the performance of these architectures and prototypes by deploying a wide range of deep learning data sets,” adds Pleros. “These data sets included image recognition, cybersecurity network traffic monitoring and optical communication.”
Commercialisation of new hardware components
The project’s successful investigation and testing of various plasmonic materials has helped to pave the way for the development of new computer components, capable of significantly outperforming electronic counterparts in terms of energy efficiency. A novel framework capable of identifying incoming network cybersecurity threats was also successfully developed. “The success of these activities has already led to the founding of two spin-off companies,” says Pleros. “The aim of these is to commercialise the material technologies we developed, such as barium titanate modulators.” This pioneering work in the fields of neuromorphic computing and AI-network cybersecurity has also led to the submission of four American patents, which could pave the way for future commercial exploitation. “A significant contribution of this project has been to highlight the feasibility of pursuing theoretical roadmaps of neuromorphic computing,” notes Pleros. “We were also able to address some of the technical and architectural challenges that have to be overcome, in order to bring this technology from the lab into real applications.”
Keywords
PlasmoniAC, neuromorphic, computing, synapses, neural, AI, plasmonic
