Learning - Growing circuits

On September 21, 2003 by Brian Taba

The genome's billion (10^9) bits cannot specify where each synapse in the brain goes—a hundred million times more bits are required to list the quadrillion (10^15) connections among the brain's trillion (10^12) neurons. The problem that computer architects worry about—how to use the trillion (10^12) transistors that fit on a silicon wafer—pales in comparison.

Having exhausted all information available in the genome, neural circuits customize themselves through internal and external interaction, a learning process known as epigenesis.

Models of epigenesis have demonstrated that the brain's feature maps can be built simply by wiring together neurons that fire together: Light- and dark-sensitive inputs from the retina wire together to produce orientation-tuned cells; left- and right-eye inputs produce depth-tuned cells; lagged and nonlagged inputs produce (motion) direction-tuned cells.

Emulating epigenesis could allow engineers to build more complex systems.

As a chip's metal wires cannot grow, we developed softwires—virtual connections that do not require point-to-point wiring. A softwire is routed to its target by a silicon growth-cone, a model of the motile structure that tows a growing axon along a chemical trail.

**Modeling epigenetic development.** Axon-terminals from randomly-activated patches of retinal cells (top-left) excite tectal cells (top-right) to release neurotropin, a chemical that diffuses to nearby locations. Migrating up the neurotropin gradient, axons from neighboring retinal cells converge on neighboring tectal cells. Coloring retinal cells (bottom-left) and tracing the colors reveals a map forming in the tectum (bottom-right). This simulation was performed by a neuromorphic chip: electrons emulate neurotropin and softwires emulate axon-migration. [Brian Taba 2003]