http://www.neurdon.com/wp-content/uploads/2009/02/feinerman_1b.jpg
216
209
Feinerman et al. Figure 1b: logic components fabricated from hippocampal neurons
Neuromorphic technology is a young field, with little in the way of established paradigms or techniques. Most of the recent related work, however, focuses on silicon implementation of neural-inspired mechanisms. Feinerman et al. buck the trend and build reliable computation devices using actual neurons.
As the authors note, neurons are exceedingly sloppy devices on their own. To borrow electrical engineering terminology, they would be well-described as "crummy" computational building blocks. Fortunately, von Neumann and other early computer scientists developed a robust theoretical infrastructure for dealing with unreliable hardware. Feinerman et al. draw on this body of work to coax cultured hippocampal neurons into behaving like diodes, threshold devices, or AND gates. The basic technique involved culturing the neurons on a precisely patterned glass carrier. The specific pattern determined the interconnects between the neurons in the circuits, allowing a high degree of control over the resulting behavior. Von Neumann error-correcting techniques ensured that the sloppy neural signaling still produced reliable behavior.
From the perspective of SyNAPSE, this work is most significant as a means for clarifying the core approach. Feinerman and colleagues perform an impressive feat, but work in a direction orthogonal to SyNAPSE. They seek to coerce reliable von Neumann computation out of biological components, where SyNAPSE aims to embrace biological-like components to develop a new kind of computation. This new breed of device will compute without the expectation of reliable signal transmission or stable wiring.
Stepping back and looking and the wider ecosystem of computation, SyNAPSE devices and von Neumann machines fill complementary roles. Von Neumann machines thrive in cases where the reliability of individual computations is vital, memory and computation can be widely separated, and a process for computing an output can be concretely defined. SyNAPSE hardware is aimed precisely at the set of computational problems complementary to the strengths of von Neumann machines: highly distributed, poorly defined, and minimal assumptions of reliability.
Nature Physics, 2008. DOI: 10.1038/nphys1099