Optical Sensor Integrated CNN for Real-Time Computational Applications

Abstract

A previously implemented CNN layer based on Nagumo neuroelectric model [11]. Hence, the cell states are a bio-inspired system model is integrated with optical sensors excitable, which results in trigger wave propagation [12]. comprised of PN type photodiodes and the design is fabricated Implementation examples of FitzHugh-Nagumo models exist in CMOS technology. The implemented test chip is measured and obtained results are presented. The sensor integration marks a agan as OTA desilgns, such as in [13]. It was shown however, new step in prototyping of this system in CMOS, which allows that an essential part of similar functionality can be realized the cell states to be programmed via light interactions, rather using custom design approach [14], which will significantly than successive state programming procedure which was utilized improve the on-chip network size.previously. It is shown that the cell input states in the network In this aper we address a CMOS integration of our can be successfully introduced via exposure of the 2D array to p p g a light source. The followed custom design methods and this custom CNN design with PN type photodiodes. The result is remarkably compatible sensor integration can lead to compact a CNN compactly equipped with pixel sensors, which offers on-chip solutions for real-time high performance applications, great potential in computational applications as a stand-alone including but not limited to certain types of image processing. system or a building block. For this purpose, another test chip in 1.5 micron CMOS technology is implemented, which I. INTRODUCTION primarily validates the functionality of the sensor integrated The neuromorphic systems in nature are ultimate references design. Built upon this step, system adaptations to different for improving system models and their design methods. The applications will be rather straightforward, with peripheral cellular nonlinear (or neural) network (CNN) theory can modifications in terms of control circuitry which addresses extensively address such models. There are previous examples the particular computational task. Along with these efforts, of this parallelism in the literature, with their implementations. this sensor integrated CNN design will remarkably boost the A particular category deals with spatiotemporally excitable real time computational performance for specific applications. models, which are implementable via autonomous CNNs [1]-[3]. The frequently highlighted applications of such systems II. THE SYSTEM MODEL include artificial locomotion [4], [5] and path optimization The dynamical system model is a resistive-coupled 2D array [6]. A vast CNN application area regards image processing, with well-known and implemented examples such as [7]-[9]. of nonlinear cells which corresponds to a specific single layer withwelommownapproachorlelted systmpls dsign is to em-CNN as depicted in Fig. 1. The dynamical state equationThe common approach for relatedl systems dlesign iS to emotie rmtecretrlto o ahntoknd ''i ploy top-down oper...