A convolutional network reads cheques in banks
According to Bottou, Bengio and LeCun of AT&T Laboratories at CVPR-97, their system for reading the amount on cheques was built into NCR's cheque processing machines, first fielded in a bank in June 1996, and had since been reading millions of cheques a month. The amount was recognised by LeNet5, a convolutional network trained together with the whole system.
Why it matters
A convolutional network became part of how banks process documents, not a demonstration: it read handwritten and printed amounts on real cheques. The paper showed how to train a whole system of several modules by gradient, rather than each module on its own.
On a mix of business and personal cheques the system read 50 percent correctly, got 1 percent wrong and rejected 49 percent to a person: for a bank not erring matters more than reading everything. LeNet5 has 401,000 connections and about 90,000 free parameters thanks to shared weights; it was trained first on 500,000 character images, then on characters cut from cheques, and finally together with the whole system through 'graph transformer networks', a generalisation of back-propagation to modules that pass graphs to one another. The authors write that independent tests by systems integrators showed it superior to other commercial systems. In Proceedings of the IEEE (November 1998) the same authors, with Patrick Haffner, add that the integrators' test was in 1995 and that since June 1996 the system had been fielded in several banks across the US. The threshold banks set is to read half the cheques with under 1 percent error, and the system was one of the first to cross it; on 646 machine-printed business cheques it read 82 percent correctly, got 1 percent wrong and rejected 17 (the previous system: 68, 1 and 31). The paper gives the volume two ways: millions of cheques a month in its introduction, millions a day in the cheque section, the conclusion and the abstract. What the record does not claim. Only the developers' own papers of 1997 and 1998 speak of the deployment, and no independent source was found, so confidence is medium. The share of all US cheques that is often quoted appears in neither.