WISARD: a neural network made of memory chips

Igor Aleksander, Thomas Stonham and Bruce Wilkie at Brunel University built WISARD, a pattern recogniser whose neurons were 32,000 random access memory chips in a single-layer net. By Aleksander in 1983 the machine took 5-10 decisions a second on 512 x 512 pixel images; by March 1984 it had been turned into an industrial product.

The patent of Aleksander, Stonham and Wilkie (priority 27 November 1981, US Patent 4,490,847 of 25 December 1984) describes groups of frame-store locations, N-tuples, mapped to discriminator stores, one per class of pattern; a television camera image is digitised into the frame store. Among envisaged uses the patent names terminals that recognise credit cards, bank notes, or people's faces or voices. By the abstract of the 1984 Sensor Review article, Computer Recognition Systems made the system an industrial product under licence from the British Technology Group. A 1987 paper speaks of networks with over a hundred thousand functions of eight inputs each, and uses both the original Brunel WISARD and the later commercial CRS WISARD. What is said about faces. The patent names faces only among its uses. In the 1983 book Artificial Vision for Robots, the chapter by Aleksander, Stonham and Wilkie names checking faces or fingerprints as a useful possibility, and Aleksander's own chapter says that experiments on WISARD had shown that smiling and unsmiling faces can be told apart or, if wanted, faces recognised whatever their expression; the visible passages give no numbers. Stonham gives his own figures in his 1986 paper, Practical Face Recognition and Verification with WISARD (Aspects of Face Processing, pp. 426-441), which sets itself both tasks, naming a person from a face and confirming a claimed identity. He there spells the name out as Wilkie, Aleksander and Stonham's Recognition Device. Only short fragments of the paper have been read, in Google Books and without page numbers: the test had 16 subjects and 16 discriminators, and each person was shown to all sixteen; images were handled at 25 a second and the training of each discriminator took under 20 seconds; and just before subsection 5.2, on expression recognition, Stonham writes that the system has been shown to function "in real time without error" within the bounds of the 16-class problem. By real time he seems to mean handling the television picture as fast as the camera supplies it, which is the camera's speed, not a measured speed of the recogniser. Vicki Bruce's 1988 summary (as reprinted in Google Books) says of WISARD applied to faces that it recognises any one of 16 individuals, with an n-tuple of four inputs and an input of 153 x 214 bits (less than the 512 x 512 that Aleksander gives for the machine), and that about 200 training images of each person are needed for a response consistently above 95 per cent of the maximum. That is a secondhand account with no source given in the fragments seen, and its percentage is a share of the maximum response, not of correct answers. A survey of face recognition by the US National Institute of Standards and Technology (NISTIR 5465, July 1994, pp. 26-27) summarises Stonham's paper in the same way, eight years on and without measuring anything: a single-layer network with one classifier for each person that, it writes, needs typically 200-400 presentations to train each classifier, the training patterns including shifts and changes of expression, with sixteen classifiers for sixteen persons; it adds that the sample is too small to conclude anything about large data sets, and gives no speed or percentage. What the record does not claim: a percentage of accuracy, the number of test presentations, or the number of memory chips for faces. The fragments read do not hold them; the full text of Stonham's paper and of the 1984 Sensor Review article (only its abstract is open) was not read, so his statement about working without error is unchecked. His figures are from a 1986 paper, so they do not show that such a recogniser ran before March 1984, and no year of the experiments appears in what was read. A 2013 Brunel thesis (Khaki) calls WISARD the first attempt at automatic face recognition in real time and says it was patented and sold commercially by 1984; it restates the literature and measures nothing. A survey by Aleksander and co-authors (2009) dates the commercial launch to 1986 without giving a source, and says UK police bought systems for face recognition and fingerprint processing; the 1984 abstract already speaks of an industrial product but gives no date of sale, and the survey has no figures.