Abstract:
In many practical applications of learning systems
to problems of pattern recognition it has been realized
and explicitly noted in the literature that linear discriminations
are inadequate. On the other hand, it has also
been noted that very little is known about the training
of non-linear systems.
A reasonable compromise between linearity and high
complexity is what is called a 'committee machine,' i.e.
a collection of linear systems each performing a linear
threshold function (subject to adaptation) with an overall
element (as the majority rule) to express the final
diagnosis.
In this paper we will present a system of algorithms
which effectively locates a committee machine which uses
majority or veto logic. The algorithms are error-correction
techniques, which in general perform as many adjustments
in training as known algorithms, but with the added feature that in some cases the algorithms will allow the machine
to misjudge some samples which are deemed to be noisy or
otherwise abnormal without implementing, in relation to
these samples, significant change in the committee members.
Experimental results are presented using artificially
generated data in 2-space, hand-printed letters A and R
(Munson), disconnected-connected 3 x 3 arrays, absence-presence
of the code 1101, and 3 x 3 quasi-randomly
generated arrays (Michalski).
