Highleyman’s Data: Pattern Recognition, Bell Labs and the Birth of Reference Datasets

Abstract This article traces the history of the 2,300 alphanumeric characters known as “Highleyman’s Data”, collected at the turn of the 1960s by Bell Labs electrical engineer Wilbur H. Highleyman. Following a brief controversy with another pattern recognition researcher, Woodrow W. Bledsoe, the samples gradually became a reference dataset used for the benchmarking of numerous character recognition methods throughout the 1960s and1970s. Looking back at Highleyman’s early career, this article first shows the influence of the synergetic organizational culture at Bell Telephone Laboratories (Murray Hill, New Jersey), where industrial projects for data transmission over telephone lines embraced the statistical principles of quality control, data analysis and operations research. By comparing the historical trajectory of Highleyman’s data with another character recognition dataset, collected by Herbert Sherman (MIT Lincoln Laboratory) during the same period, this genealogy then demonstrates that the growing popularity of the former stemmed less from the quantity and quality of the alphanumeric characters than from the metrological possibility of evaluating several methods on the same samples. Finally, Highleyman’s offer to make available what is considered by some to be the first machine learning benchmark preceded a collective initiative to share reference datasets through the Technical Committee on Pattern Recognition affiliated to the Computer Group (Institute of Electrical and Electronics Engineers).