Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP Path: utzoo!mnetor!uunet!husc6!mit-eddie!uw-beaver!cornell!batcomputer!pyramid!hplabs!well!wcalvin From: wcalvin@well.UUCP (William Calvin) Newsgroups: comp.ai.neural-nets Subject: Contest to rename non-neural "Neural Networks" Message-ID: <4210@well.UUCP> Date: Wed, 14-Oct-87 01:57:58 EDT Article-I.D.: well.4210 Posted: Wed Oct 14 01:57:58 1987 Date-Received: Thu, 15-Oct-87 23:36:53 EDT Lines: 75 Keywords: neurobiology, retina, distributed, plastic A most interesting new field is unfortunately named: neural networks. They are, of course, comprised of real neurons with real DNA, real developmental histories, etc. Neurophysiologists have studied real neural networks since 1929, though without understanding circuit principles until about 1960 or so when the analysis of lateral inhibition in the eye of the horseshoe crab _Limulus_ showed us how circuits could perform inverse transforms to recover what poor optics had smeared out. And neural networks are an important field of specialization within modern neurobiology: we are taking collections of neurons called ganglia, such as the 30-cell crustacean stomatogastric ganglion, and getting to know each of their neurons as an individual, mapping the synaptic connections between neurons, and so beginning to understand how circuit properties emerge (for a nice full-wave rectifier, see Graubard and Hartline in the 7/31 issue of _Science_). The lobster ganglion produces two entirely different rhythms simultaneously, and both can be substantially altered by background biases such as neurohormone levels, some of which virtually "rewire" the circuit. One way we understand emergent properties is to simulate those networks -- and in a way far more sophisticated than so-called "neural networks" in AI. For example, each cell's tranfer function is somewhat different: the free parameters in the simulation are minimized by experimentally determining each cell's time-dependent response to inputs, and each synaptic interconnection's changing strength with repeated use. See Dan Hartline's chapter in THE CRUSTACEAN STOMATOGASTRIC SYSTEM, edited by Selverston and Moulins (Springer Verlag 1987), for the state of the physiological art in simulating real neural networks. But it seems absurd for neurobiologists to have to start talking about "real neural networks" just because the AI folk didn't learn their lessons. And I'm not referring to ignorance of neurobiology, though that too is a sore point: remember the hyperbole in the old days when every digital computer got called a "brain"? And how soon no self-respecting computer person would call a computer a brain for fear of being thought a beginner? SO why are we now seeing this nonsense of calling any plastic network of pseudo-neurons a "neural network"? For some simulations, it seems appropriate to use "neural network" in referring to the computer model: those simulations of lobster networks, the simulations of the retina using state-of-the-neurobiological-art parameters, etc. But most so-called "neural networks" in AI don't even have the ambition to simulate a real neural circuit: they are seeking shortcuts around formal programming, a plastic network that can be shaped up by training until it performs a desired task (and then perhaps cloned). Particularly when stochastic sequencing is implemented in neural-like nets, we are going to see some strikingly humanlike capabilities emerge (see my article "The brain as a Darwin Machine" shortly to appear in _Nature_). So how about a contest to devise a new name for this wonderful new field that will give it an identify respectful of, but independent from, the endeavors concerning real neural nets? Some possibilities: Pseudo-neural networks Neuroid networks Neural-like networks Plastic networks Parallel Distributed Nets Cellular Networks Dry Nets Perhaps a look-alike, the way Dawkins coined "meme" as the cultural equivalent of "gene"? "Seural" as a silicon version of "neural"? Yes, I know, they don't trip alliteratively off the tongue like neural nets. But that phrase is already taken, has been for a quarter century, and constitutes an active field that modellers ought to be mining-for- leads rather than regularly re-inventing the wheel. William H. Calvin University of Washington NJ-15 Seattle WA 98195 USA wcalvin@well.uucp wcalvin@uwalocke.bitnet 206/328-1192