So what is 'Neuromorphic' anyway?
This term is intended to describe the design of computational devices that incorporate the physical structure of neurons and neural circuits to solve the difficult computational problems that biological systems solve in real-time (e.g., vision and audition) and with superb power efficiency. Neuromorphic VLSI is therefore the implementation of this computational device using the medium of modern silicon fabrication processes. ‘Biomorphic Robotics’, likewise, is the design of robotic systems that are shaped like biological examples (e.g., a strong spring-like, kangaroo-inspired, foot-tendon system for robots that bounce). These two concepts are inextricably linked because evolutionary processes sculpt both brain and body to work with each other.
Electronic modeling of neural systems and robotic modeling of behavior both have a long history of testing and demonstrating our current understanding of biological systems. In most cases, however, these two endeavors have been treated separately either due to a lack of knowledge of the underlying neural control at the cellular and system levels (neuroscience), limitations of electronic realizations (fabrication technology), shortcomings of available robotic technology (materials, actuators, batteries), or the limited appreciation of the critical synergy between brain and body. In the last decade or so, with the continued advances in digital computing, rapid advances in neuroscience, and the latest consumer demand for portable electronics, many new technological opportunities for biologically-inspired robotics are emerging. While early efforts at modeling neural circuits with discrete components provided limited insight due to the small numbers of cells, the development of a toolbox of analog VLSI primitives for describing neural structures has launched a field of engineering known as Neuromorphic VLSI. This approach involves designing hybrid analog/digital VLSI circuits that mirror neural algorithms in both signal representation and relevant morphology. Using commercial silicon foundries, large arrays of analog and digital circuits can be fabricated on a single chip relatively inexpensively to perform the massively-parallel signal processing known to occur in the retina, the cochlea, the midbrain, the cortex, the spinal cord, and many other sensorimotor structures.
By utilizing dedicated parallel analog circuits, low-precision computations can be performed in real-time (or faster than real-time) with many orders of magnitude less power than on a general-purpose computer. While the difficulty of translating these devices into useful neural modeling tools remains, the speed of computation and the small physical size promise to allow real-time modeling of complex sensorimotor interactions that were previously impossible (e.g. echolocation-based flight control of a micro aerial vehicle). VLSI-based neural models also stand out as useful computational tools where a large range of time constants is desired such as in spike-based synaptic learning rules or in spike-based models of motor control, where real-time neuron-to-world interactivity is desired.