摘要 |
<p>1,002,405. Learning networks. STANDARD TELEPHONES & CABLES Ltd. Aug. 2, 1963 [Aug. 8, 1962], No. 30725/63. Addition to 939,134. Heading G4R. In a learning matrix of the kind described in the parent Specification, which comprises an array of cores, the cores of each row being set in a learned pattern, the binary pattern to be recognized is applied in the form of radio frequency currents of one of two phases displaced by 180 degrees, all cores also receiving a second radio frequency current, the match of each stored pattern with the input pattern being represented as a voltage amplitude of a frequency which is the sum or difference of the first and second frequencies. Generator G1 is connected to a split-secondary transformer giving voltages having the same amplitudes but opposite phases T(W1) and T(W1) + 180 degrees. Switches S, which may each be a pair of gates controlled by a trigger, apply the input pattern and each core generates in the corresponding row lead a voltage in phase or 180 degrees out of phase with the applied current according to the state of the core. The best match row will produce a maximum signal having a frequency equal to the difference between the frequencies of the first input current i(# 1 ) and the second i(# 2 ). The worst match also produces a maximum signal but of opposite phase. The best and worst match signals can be distinguished by phase discrimination, e.g. using a ring modulator. In the form of Fig. 1, however, the antiphase signal is cancelled by adding a voltage UZ derived by combining signal i(# 1 and i(# 2 ) in core V. The useful signals representing the extent of the match are separated by selective L-C networks SN and applied to a maximum signal detector.</p> |