MINSKY@OZ.AI.MIT.EDU (10/24/87)
Terms like "neural networks" were in general use in the 1940's. To see its various forms I suggest looking through the Bulletin of Mathematical Biophysics in those years. For example, there is a 1943 paper by Landahl, McCulloch and Pitts called "A statistical consequence of the logical calculus of Nervous Nets" and a 1945 paper by McCulloch and Pitts called "A heterarchy of values determined by the topology of Nervous Nets. It is true that Papert and I confused this with the title of another McCulloch Pitts 1943 paper, which used the term "nervous activity" instead. Both papers were published together in the same journal issue. In any case, "neural networks" and "nervous nets" were already in the current jargon. In the original of my 1954 thesis, I called them "Neural-Analog Networks, evidently being a little cautious. But in the same year I retitled it for publication (for University Microfilms) as "Neural Nets and the Brain Model Problem". My own copy has "Neural Netorks and the ..." printed on its cover. My recollection is that we all called them, simply, "neural nets". A paper of Leo Verbeek has "neuronal nets" in its title; a paper of Grey Walter used "Networks of Neurons"; Ashby had a 1950 paper about "randomly assembled nerve networks. Farley and Clark wrote about "networks of neuron-like elements". S.C.Kleene's great 1956 paper on regular expressions was entitled "Representation of events in Nerve Nets and Finite Automata". Should we continue to use the term? As Korzybski said, the map is not the world. When a neurologist invents a theory of how brains learn, and calls THAT a neural network, and complains that other theories are not entitled to use that word, well, there is a problem. For even a "correct" theory would apply only to some certain type of neural network. Probably we shall eventually find that there are many different kinds of biological neurons. Some of them, no doubt, will behave functionally very much like AND gates and OR gates; others will behave like McCulloch-Pitts linear threshold units; yet others will work very much like Rosenblatt's simplest perceptrons; others will participate in various other forms of back-propagated reinforcement, e.g., Hebb synapses; and so forth. In any case we need a generic term for all this. One might prefer one like "connectionist network" that does not appear to assert that we know the final truth about neurons. But I don't see that as an emergency, and "connectionist" seems too cumbersome. (Incidentally, we used to call them "connectionistic" - and that has condensed to "connectionist" for short.)