PERFECCIONAMIENTOS INTRODUCIDOS EN UN SISTEMA DE TRATAMIENTODE INFORMACION.

摘要

1457338 Adaptive processing systems NESTOR ASSOCIATES 7 Nov 1974 [6 June 1974] 48232/74 Heading G4R The system has a plurality of input terminals and one or more output terminals connected to the input terminals by a plurality of transfer units. Each transfer unit connects one of the input terminals to the, or one of the, output terminals. A signal s j on input terminal j appears at the output of the relevant transfer unit as a signal s j A ij . During a learning mode the transfer function A ij of the transfer unit is modified in accordance with the function of the product of the input signals s and the signals r i on the output terminals, such that #A ij =f(s 1 , s 2 ... s j - ...s N ; r 1 , r 2 ... r j ... r n ). It is not necessary with this system to apply desired response signals during learning. The basic arrangement is shown in Fig. 2 : each input signal s feeds a corresponding transfer unit i to provide signals s1. Summing unit #produces the sum r of signals s1 which is fed back during the learning mode to modify the transfer function A of each transfer unit such that #A ij = f(s j . r;). Instead of being the sum of signals s1, output r may be some other function of all s1. If there is a plurality of outputs r each input s may be connected to several transfer units (Fig. 3, not shown) to form an adaptive module. Modules may be connected in series, parallel, or series-parallel. The system may be used in a passive learning mode, in which it develops its own responses r 1 -r n for a given set of input signals s 1 -s n , or in an active learning mode in which it is forced to develop desired response signals for a given set of inputs by applying desired response signals to the output terminals. The output signals may be passed to threshold units so arranged that, for example, the output of each summing unit # is suppressed except for the highest output, this arrangement training the system to provide only one non-zero output signal in response to a given set of input signals. It may be desirable to allow the transfer functions to decay at a predetermined rate during learning so that details of the module's previous learning experience are rendered less significant, i.e. are "forgotten", allowing the module to generalize more rapidly. A specific embodiment of a transfer unit is described (Fig. 14, not shown).