By David Colella

It is early in the 21^st century. Robots swarm around you, negotiating their environment with such fluidity and precision that, were it not for their outward appearance, you might believe they were living animals. A miniature flying robot, maneuvering much like an insect, arrives and singles you out from a multitude of people by tracking your scent and identifying your facial features. It lands lightly in your hand to deliver a specimen of its discovery: a small sample of dirt contaminated with a leaked chemical.

Such a scenario, until recently the stuff of science fiction, is on a direct path to becoming scientific reality. This reality will be made possible through the efforts of forward-looking researchers who today are focusing on understanding biological processing mechanisms and incorporating basic principles of these mechanisms into computer-based computational engines.

Research in this area is driven in part by the desire to develop autonomous sensor and robotic systems that can interact with complex dynamic environments in a robust way, much as people and animals do. This is an achievement that has remained infeasible using current technological approaches.

One principal focus for this research is the development of novel processing paradigms in which vast arrays of neuron-like processing elements are embedded within the silicon architecture of a chip and then interconnected by a sophisticated feedback network. These "silicon neurons" are designed to process information in a way that directly mimics the processing that is found in biological neurons. Furthermore, they employ the same physical resources to perform computation: electric currents and voltage differences. In several cases, this research approach has even enabled the construction of direct links between the processing elements within a computer system and living biological neurons.

A New Paradigm for Processing Information

Silicon neurons are processing elements that are modeled directly on the structures found in biological neurons. They are manufactured so that the fundamental processing principles of biological cells are emulated within the hardware architectures. These designs therefore provide a novel approach for processing information. They are markedly distinct from well-known earlier attempts at modeling biological neurons (as in artificial neural networks) in both fidelity and in how information is represented and processed. This new paradigm provides a number of advantages, such as reduced power requirements and smaller size.

The greatest advantage of silicon neurons, however, is their increased processing capabilities. Networks of silicon neurons exhibit significantly greater operational dynamics, which translates into their being able to capture and process a much more diverse collection of information. This is true even for networks with a small number of silicon neurons. Furthermore, these networks have been shown to exhibit characteristics that are typical for biological processing but that had not been reproduced using processing techniques currently in use.

Modeling Dendrites

The attempt to base processing elements on biological models is nothing new. The first artificial neuron model, the perceptron of McCulloch and Pitts, dates back to the mid-1940s. But silicon neurons are novel in that they embed the model structures directly in the hardware architecture of the computer chip. As opposed to computing with an algorithm or set of rules (which typical artificial neural networks do), the silicon neurons perform computation using the physics of the hardware in which they are embedded—an exact parallel to how computation is performed in biological cells. For example, both biological neurons and silicon neurons achieve signaling using electric currents and voltages. Signaling between elements is enhanced or inhibited through "gates" across voltage drops and resistances.

The improvement afforded by silicon neurons is a direct result of the increased complexity of their basic structure. This complexity is partly due to the greater emphasis researchers now place on incorporating the activity present in the dendrites of biological neurons, the primary receiving stations for the output of other neurons.

As is the case with biological neurons, silicon neurons contain a tree-like structure of dendrites that interweaves, and complicates, the communication flow between neurons. The dendrite structures are modeled as a series of individual compartments, each having the ability to receive different types of input signals (such as excitatory and inhibitory) from any or all other neurons, including itself. The spatial distribution of the dendrite compartments provides many desirable processing characteristics, including spatial and temporal dispersion of input signals. It is believed that such characteristics play a vital role in processing within biological neurons.

Complementary Efforts Advance Research

Several groups lead the field in research in this area. The most prominent are the California Institute of Technology and the University of Delaware in the United States, the Institute of Neuroinformatics in Switzerland, and the Max Planck Institute for Biochemistry in Germany.

The Engineering Research Center at Caltech (http://www.erc.caltech.edu) has an extensive program to model biological functionality for application use and has cultivated a multitude of young researchers in this field. The University of Delaware (http://www.ece.udel.edu/~elias/neuromorphicSystems/index.html) is the leader in developing connectivity mechanisms for silicon neurons. Its staff has developed a network of "virtual wires" that enables interaction between the various compartments of all the neurons. (See illustration below.)

The Institute of Neuroinformatics (http://www.ini.unizh.ch) is breaking ground in the study of the operational dynamics for small networks of silicon neurons. In one recent study, their researchers developed a silicon neuron network that exhibited dual processing properties found in biological neurons that, until now, had been considered incompatible for the purposes of standard processing mechanisms. The Institute has also been particularly successful in fostering collaborative efforts in this field with other institutions, such as Bell Laboratories and MIT Lincoln Laboratories. The Max Planck Institute (http://www.biochem.mpg.de/mnphys) has focused on developing interaction between living biological neurons and computer hardware. They recently verified the passing of information through a biological neuron from one computer gate to another.

The two major funding sources for research on silicon neurons in the United States are the National Science Foundation and the Office of Naval Research. Both organizations support programs focused on applying biological processing mechanisms and biological functionality to engineering systems. The silicon neuron work is but a part of these extensive programs.

Clearing the Hurdles

Many roadblocks must still be hurdled before research on silicon neurons has the chance to provide the sought-after processing capabilities. One major hurdle is understanding the impact of model fidelity. It is not known a priori what level of model fidelity is required to ensure particular functionality. Whereas for one application a relatively simple model is sufficient, a different application could well require more complex neuron structures and more sophisticated network connectivity.

A second major unresolved issue is how best to "train" a network of silicon neurons for each particular application. In other words, a well-defined methodology for learning is needed. Currently, all approaches for devising network architectures and connectivity for any given problem remain ad hoc and often include a significant expense for trial and error.

A Model for Success?

Biology can provide important lessons for the development of the processing engines that drive many of our engineering systems. Attempts to incorporate fundamental principles and structural foundations from biology within our engineering systems (thus leading to the "smart" robots we met earlier) require a deeper understanding of the biological mechanisms we are trying to emulate. Our success will likely depend on how well we are able to translate knowledge of biological processes into computational mechanisms that mirror similar mechanisms found in biology. The silicon neuron, a computational device for processing information that is modeled on its biological counterpart, should play a major role in accomplishing this migration.

---

For more information, please contact David Colella using the employee directory.

---