The Saga of Highleyman's Data.
The first machine learning benchmark dates back to the late 1950s. Few used it and even fewer still remembered it by the time benchmarks became widely used in machine learning in the late 1980s.
In 1959 at Bell Labs, Bill Highleyman and Louis Kamenstky designed a scanner to evaluate character recognition techniques. Their goal was “to facilitate a systematic study of character-recognition techniques and an evaluation of methods prior to actual machine development.” It was not clear at the time which part of the computations should be done in special purpose hardware and which parts should be done with more general computers. Highleyman later patented an OCR scheme that we recognize today as a convolutional neural network with convolutions optically computed as part of the scanning.
Highleyman and Kamentsky used their scanner to create a dataset of 1800 alphanumeric characters. They gathered the 26 letters of the alphabet and 10 digits from 50 different writers. Each character in their corpus was scanned in binary at a resolution of 12 x 12 and stored on punch cards that were compatible with the IBM 704, the GPGPU of the era.
_social_media.jpg)
