TEEMAL, an adaptive training procedure for a two layer system of linear threshold elements

Graduation date: 1973In many practical applications of learning systems\ud to problems of pattern recognition it has been realized\ud and explicitly noted in the literature that linear discriminations\ud are inadequate. On the other hand, it has also\ud been noted that very little is known about the training\ud of non-linear systems.\ud A reasonable compromise between linearity and high\ud complexity is what is called a 'committee machine,' i.e.\ud a collection of linear systems each performing a linear\ud threshold function (subject to adaptation) with an overall\ud element (as the majority rule) to express the final\ud diagnosis.\ud In this paper we will present a system of algorithms\ud which effectively locates a committee machine which uses\ud majority or veto logic. The algorithms are error-correction\ud techniques, which in general perform as many adjustments\ud in training as known algorithms, but with the added feature that in some cases the algorithms will allow the machine\ud to misjudge some samples which are deemed to be noisy or\ud otherwise abnormal without implementing, in relation to\ud these samples, significant change in the committee members.\ud Experimental results are presented using artificially\ud generated data in 2-space, hand-printed letters A and R\ud (Munson), disconnected-connected 3 x 3 arrays, absence-presence\ud of the code 1101, and 3 x 3 quasi-randomly\ud generated arrays (Michalski)