Future Memcomputing Arrays for Next Generation Computer Vision

Lead Research Organisation: University of Hull
Department Name: Physics

Abstract

Real-time processing of high-resolution images is essential for many new A.I. technologies. However, currently the computational needs, cost and energy requirements are prohibitive for many mainstream applications, not to mention the lack of portability of the processing systems or latency issues if computing is done via the cloud. New approaches must be adopted to meet these challenging demands.

Highly parallel computing is widely agreed as the only viable way to achieve the level of performance needed for real-time imaging. However, the complexity and number of circuit components required to achieve this with traditional semiconductor CMOS approaches impacts the overall system's speed and optical resolution. Thus, there is a need to develop new types of circuit components that are specifically designed for neuromorphic computing.

Memristors are two terminal electronic devices that have attracted intense research interest owing to their simple fabrication, low-cost manufacture, low power operation and their capacity for ultra-high density, non-volatile data storage. In recent years, memristor performances have advanced considerably. Very high levels of endurance (120 billion cycles) and retention (>10 years) have been achieved, and ultra-high-density cross-bar arrays have been realized with scalability down to 2 nm. However, it is their ability to emulate the memory and learning properties of biological synapses and their potential to produce a new generation of ultra-high performance artificial intelligent devices that has ignited researchers' interest in these remarkable devices. Many basic neuronal functions have been demonstrated and memristor arrays have been shown to efficiently carry out processing in the analogue domain, removing the computational bottlenecks associated with the large number of vector-matrix operations. This combined with recent improvements in device reliability gives a promising outlook for their future use as the world seeks new technologies to circumvent the end of Moore's law and the problems of traditional von Neumann computing, which has inherent bottlenecks in the way information is processed and transported.

Recently there has been a drive towards the development of memristor devices that can be read, written or have their switching characteristics modified by the application of light. The development of these devices, termed Optical Memristors, arises due to several potential benefits. Optical systems are free of sources of electronic noise and capacitive coupling effects, which limit the operating speed of traditional electronic devices. The combination of memristor technology with optical systems offers the additional advantage of high-speed data routing while consuming little power, as well as integration as a building block within future optical computer architectures.

In this proposal a new type of computer vision recognition system is proposed based on optical memristors (OM) and cellular nonlinear networks (CNN) that leverages the unique capacity of OM's to detect light and store information while also exploiting CNN's ability to simultaneously process the information in all cells at once. This will enable ultra-fast real-time (in-memory and parallel) computation. The approach outlined contrasts with standard vision recognition systems which are inherently limited by data transfer bottlenecks and the slow, serial processing of information. This research will therefore pave the way to a new generation of ultra-fast, high-resolution vision recognition systems that will impact a wide range of current societal needs (e.g. safer autonomous driving, better security systems) and numerous applications in medicine (e.g. high throughput cell imaging for early cancer diagnostics).

Publications

10 25 50