Neuromorphic Engineering

Dr. Kwabena Boahen and colleagues at the University of Pennsylvania are performing research which combines electical engineering with neurobiology. They are attempting to reverse engineer biological nervous systems, and then to recreate the circuits they find in silicon. This technique of building biologically similar electronic circuits was first pioneered by Professor Carver Mead of Caltech in the late 1980's. He named the technique "Neuromorphic Engineering".

The long term aim of Dr. Boahen's laboratory is to build complete artificial brains. Of course they are still a long way from this. They have, however, already been successful in recreating a small part of a brain, namely the neural circuits of the eye's retina. They are also making progress in developing microprocessors whose internal wiring is able to "grow" and learn.

Morphing the Retina:
The retina of the eye is able to encode information 1000x more efficiently than current video cameras. In order to try to understand how the retina achieves this, the inner and outer retinal micro-circuits were reverse engineered. These circuits were then morphed into VLSI CMOS to implement parallel visual pathways on a chip.

This chip, called Visio1, models the four major ganglion cell types. It has 104x96 photoreceptors, 4x52x48 ganglion cells, and a die size of 9.25x9.67mm2. It consumes just 11.5mW when running at 5 spikes/second/neuron. Networks of these various ganglion cells are able to detect edges and distinguish directions of motion.

It is hoped that with advances in CMOS technology the laboratory will be able to model five different classes of ganglion cells, each with half a dozen synapses. More transistors will be crammed onto each chip, and the overall chip size reduced. This would allow these retinomorphic chips to be used in visual prostheses and to restore vision to patients who are blind due to damaged photoreceptors.

Learning by Growing:
It is of course obvious that the brain cannot be programmed in the same way that today's computers are. Instead the brain dynamically reprograms itself. The team in Boahen's lab is also trying to recreate this behaviour in silicon.

From their website: "As a chip’s metal wires cannot grow, we have developed virtual connections that can be reconfigured dynamically. These virtual axons connect neuromorphs together without using a dedicated point-to-point wire. And they can be attracted toward appropriate targets by sensing local activity mimicking the activity-dependent release of rapidly diffusible agents like nitric oxide. By recreating these activity-dependent developmental processes in silicon, we hope to sculpt a structure as variegated, as versatile, and as efficient as the brain."

Links:
University of Pennsylvania: www.neuroengineering.upenn.edu