Neuromorphic Systems

Abstract

Neuromorphic systems is the field of engineering that attempts to imitate the operation and structure of biologi-cal sensing and information processing nervous systems. The word " neuromorphic, " originally introduced by Carver Mead (1, 2), is understood here as something that attempts to artificially recreate biological nervous systems. The way simple living beings sense and process information so very efficiently, relying on slow and often defective components called " neurons " and yet consuming so very little power, is truly amazing. Human beings have no difficulty in recog-nizing sophisticated objects within complex moving scenes and commanding our arms and fingers to grasp them naturally and effortlessly. However, we have still not yet succeeded in building machines that can do similar tasks at speeds and with power consumption similar to those of their biological counterparts. In this article, our aim is to provide an overview of the core elements of artificial neuromorphic systems as they are understood today (although we realize that the result will almost inevitably be incomplete). First, neuromorphic systems can be divided into two main (and not necessarily overlap-free) categories: " sensory subsystems " (corre-sponding to biological organs such as eyes, ears, noses, and skin) and " processing subsystems " (corresponding to the biological brain). But neuromorphic systems must also take into account three separate, complementary pro-cesses: (a) information computation, (b) information com-munication, and (c) information storage. These three processes are habitually performed in standard man-made computers, but there they are typically implemented on separate, physical " pieces of hardware " (either chips or circuits within chips). In neuromorphic systems, such separation is not so clear-cut and all three processes are often implemented in one and the same physical element over and over again. For example, a synapse connecting two neurons performs computing, communication, and storage tasks. In biological neural systems, information is typically exchanged through electrical spikes, except when some local tissue is used such as the retina, where continuous signals transmit information to neighboring cells. Here we will focus mainly on artificial neuromorphic systems that encode, transmit, and process information in the form of spikes. We will use the term " events " to reference such spikes. In general, the events discussed will be digital electronic multibit signals capable of traveling very quickly (typically, in fractions of microseconds) between electronic components. Artificial systems of this type are commonly known as " event-driven neuromorphic systems " (3), and their event-driven intercommunication strategy is called " address event representation " (AER) (4–9). In the following sections, we will very briefly describe some of the principles involved (components, communica-tions, learning) and look at some examples of sensors (retinas, cochleae, skins noses) and event computing sys-tems. Sources or material will also be referenced for further reading.