Artificial Intelligence has two sides: mimicking human intelligence by performing similar tasks, and modeling the natural structures and functions. This article appears under a banner of “Artificial Intelligence,” probably because it describes nerve cells and some analogs, and perhaps because it leaps to speculation about a computer structure based on components with similarities to nerve cells. As a technical article, it should be dismissed out of hand on the evidence of the title alone. But in BYTE it is most likely to reach a number of technical neophytes whose imaginations will be stirred by this speculative, popularized article.

Most of the article describes the major electrical, chemical, and geometrical processes of two classes of nerve cells: transmitting or “output cells,” and processing “interneuron” cells. Elementary but clear descriptions are interspersed with appeals to analogies in electronics which may not always be illuminating to the naive reader. But if you don’t mind your kidney being compared to a battery charger, then the descriptions of actions of “Active Axons,” “Driven Dendrites,” and supporting structures will be easy to accept. Although I am not qualified to rule on the accuracy, I found the descriptions generally consistent. When Stevens says an output neuron’s output is a “fixed waveform” I assume he means a nearly constant pulse shape. When he says, “. . .interneurons have graded analog inputs but also have graded analog outputs,” it seems he refers to variable-amplitude pulses.

A major part of the article lays the basis for the premise in the title. This part describes an electrical-circuit model for the electrodynamics of the passive components of a single cell. The dendrite is said to be modeled by a battery-powered resistor-capacitor ladder circuit. The action of a synapse input is said to be modeled by a “gate” consisting of a changing conductance in series with one battery. Although this section contains one of the highlights of the paper (a diagram of an axon, not a dendrite, constructed from a sequence of scanning electron micrographs of thin sections), the section lacks a clear comparison of the pulses produced by the model with some of real nerves. The results of some simulation calculations are shown, but in the graph it seemed quite unclear just what was being plotted.

Stevens goes on to discuss nerve circuits as if there are just two significantly different behaviors to model. He then sketches part of an electrical-circuit analog, and presents as examples only cells from the retina. Yet in the last part of the paper, he leaps to the statement that one can, by etching shapes of cells into silicon, “quite easily” create silly-sounding “silicrons” which “could simulate brain circuits.” There is no intervening discussion of any variety in brain neurons or of a role which might be played by major modular structures. Since the gross functions of circuits with large numbers of nerve cells (i.e., optic bundles or brains) and “how the brain modifies its own circuits” are not well understood, the time at which one could build a computer “using circuitry copied directly from the brain. . .continues to be ‘Probably not right away.”’ What Stevens leaves unaddressed is why we would want to, except for the understanding gained by modeling. He compares the complexity of structure of an integrated circuit as much lower than that of a retina and confuses that with “performance.” He implies that because simulating the actions of the respective nerves would require an enormous amount of processing, a computer built on the basis of a neurological model is a candidate for a “sixth-generation.”

Despite the omissions of fact and logic, here is one vote to encourage such imaginative pieces. The article is likely to stimulate any reader: it contains some significant illustrations, should please the dilettante, and contains material for some good fun.