BY D. M. MACKAY Wheatstone Physics Laboratory, King's College, London
The past decade or two have seen a great variety of attempts to simulate the nervous system, usually by models drawn from the field of electronic circuits. At one extreme we have had the digital computor, discrete in its structure and discrete in its symbolic variables (which are often reduced to 'all' or 'none'). For the firing of a single peripheral nerve-cell, with its all-or-none character, the analogy was tempting; and there seems to be little doubt that 'coincidence detection' after the manner of digital circuitry does take place where several synapses converge on single neurons ( Lorente ed Nó, 1939). We owe to McCulloch & Pitts ( 1943, 1947) the demonstration that theoretical 'neurons' on this principle could be combined to imitate any specifiable type of information processing in living organisms; an 'existence theorem' which rapidly encouraged (or provoked) a mounting stream of variants and alternatives.
For afferent and efferent systems in particular, it has become clear that the thought-forms of 'analogue' rather than digital computing are more appropriate. Frequency-modulation, rather than binary coding, seems to be the rule. To use the interval between nerve-impulses as a continuous 'analogue' variable need in fact be no less efficient than to use their presence or absence as a digital variable ( MacKay & McCulloch, 1952, 1953); and 'analogue' servo-models have thrown much light, for example, on the mechanisms of control of skeletal muscle ( Merton, 1953, Eldredet al. 1953).
We thus arrive at a second, more general, class of model--the 'information-flow model'--in which discrete pathways are invested with signals that may be effectively continuous. As I have argued elsewhere ( MacKay 1951a, 1954, 1956a) these pathways should comprise not only conventional (presumably neural) signal-channels, but also channels (which might well be biochemical) through which the probabilities of neural function are modified). For the broad outlines of large-scale organization of the C.N.S. one may hope that this type of model will have permanent value, especially in clinical work (from which, incidentally, it may be expected to gain much of its structure) ( MacKay, 1954).
The object of the present brief paper, however, is to explore a further step. When enough cross-connexions are formed in a network of discrete pathways, the characteristic intuitions of the servo-engineer begin to fail him. The concepts of feedback, and even of input and output, become ambiguous and