Neural Computation Competition

[John J. Hopfield]

ANNOUNCING A NEURAL COMPUTATION COMPETITION
-------------------------------------------

by  John Hopfield (hopfield at princeton.edu)
and Carlos Brody  (carlos at cns.nyu.edu)

Web Site (with preprint):  http://neuron.princeton.edu/~moment

How do we identify spoken words? How do we recognize textures with our
fingers? How do we integrate information over time to perceive
structure-from-motion? All of the above are examples of the capacity
of nervous systems to integrate information over time and recognize
spatiotemporal patterns. How this is actually done in the brain
remains, of course, a relative mystery.

Here we wish to introduce a novel neural network, built on novel
neurocomputational principles, which recognizes spatiotemporal
patterns. We exemplify this with robust recognition of a monosyllabic
utterance.

Unusually, we have chosen to present this novel network in the form of
a competition. We have treated the network as if it were a real
biological organism, and we present only the results of the
"experiments" that one would normally carry out with such a
creature. (We have playfully dubbed the network "mus silicium, the
sonoran desert sand mouse.") The experimental results presented are,
we believe, sufficient to deduce the novel principles on which the
network operates. For people wishing to carry out further experiments,
we have constructed an interactive web site. Sound stimuli may be
uploaded to this web site, and the resulting spike train responses of
the mus silicium's neurons may be downloaded from the site.

THE CHALLENGE, THEN, AND THE GOAL OF THE COMPETITION, IS TO DISCOVER,
BASED ON THE "EXPERIMENTAL" RESULTS, HOW THE NETWORK OPERATES. In this
the competition exactly mimics wet neurobiology. The rules of the
competition are detailed further below and on the web site; on
Dec. 14th, we will anounce the results of the competition and reveal
the principles of the network's operation.

However, although we have formalized the competition, and we provide
specific rules and prizes, we do wish to emphasize that the greatest
benefits of the exercise are to be found simply by *trying to DEDUCE,
based on the extant "experimental" results, how the network
operates*. The prizes and formal competition are there merely for
extra fun, they contain no scientific information in themselves.

The interactive web site/homepage for this network and associated
competition, containing full information on the network, a preprint,
and the competition rules, is:

          http://neuron.princeton.edu/~moment.

To be considered, competition entries must be received by
carlos at cns.nyu.edu or hopfield at princeton.edu PRIOR to Friday 1st
December 2000.

                       ----------
                       MOTIVATION
                       ----------

We wish to note that there are two quite distinct aspects to what we
are doing:

(1) We are presenting novel computational principles which we believe
    will be of great interest to the computational neuroscience and
    connectionist community. Because of the method of presentation, we
    cannot demonstrate these principles to you directly. At this
    point, we can only give you our word that, in our own estimation,
    they are simple, powerful, biologically plausible, and novel.

(2) In computational neuroscience modeling one typically
    "guesstimates", on the basis of experience, which aspects of the
    data are relevant and which are irrelevant. One then builds a
    model consistent with the chosen relevant aspects.  This process
    is an example of what we will term "RATIONAL" reasoning. The
    rational view as applied to neuroscience allows data properties to
    be caused by unknown factors such as unobserved cell types, novel
    channels, etc. This approach should be distinguished from
    DEDUCTIVE reasoning, for which once basic properties are
    sufficiently known, new conclusions can be logically deduced from
    even apparently minor aspects of inconsistency or the
    unexpected. Deduction is enormously different from even the most
    rational "guesstimation." Given the quantity of diverse data now
    available in neuroscience, the question we are posing with our
    network and which we underscore through the competition is: to
    what extent is a deductive approach becoming useful in
    neurobiology?

                 ----------------
                 THE COMPETITIONS
                 ----------------

Further details of the competition rules can be found through the main
site, http://neuron.princeton.edu/~moment.

There are two "mus silicium" competitions. You can enter one
or both, as you wish.

The goal of CONTEST A is to write a 750 word essay, with or without
equations, describing how you convincingly reasoned deductively from
the experimental results presented in the preprint (augmented if
desired with further experiments you may have carried out or seen on
the web site) to understand the principles by which the "mus
silicium" network operates.  The cash prize for the best essay is
$500, for the runner-up $200.

The goal of CONTEST B is to generate an artificial neural network, of
approximately the same size, complexity, and biological plausibility
as "mus silicium", which recognizes the spoken word "one". The actual
"mus silicium" network can be thought of as merely one possible entry
meting the specifications of this contest. The cash prize for the best
recognizer network is $500, for the runner-up $200. On the website, we
detail what we mean by a network of "approximately same size,
complexity", etc. as "mus silicium." For Contest B, we provide on the
web site a front end that turns sound files into spatiotemporal
patterns of timed events. Thus, Contest B is essentially about
generating a network that recognizes spatiotemporal patterns (which
underscores the generality of the problem being addressed), while
using a limited number of simple, biologically plausible spiking
neurons. (Needless to say, we think that, if entered, "mus silicium"
would win Contest B-- but maybe we are wrong!)

For both contests, entries must be submitted by email to either
hopfield at princeton.edu or carlos at cns.nyu.edu before Dec. 1st, 2000.
Multiple author submissions are welcome.  A group of N authors can
submit N entries.