EP12: Searching for Zhang Wei: An Untold Story of Convolutional Neural Networks

Summary

This article is a correction and supplement to earlier content, focusing on Zhang Wei, a key yet overlooked figure in the early development of convolutional neural networks (CNNs). It clarifies the publication timeline of Yann LeCun's seminal CNN paper: submitted in July 1989, accepted in September, and published in December, spanning about five months. More importantly, it details Zhang Wei's contributions: at the University of Chicago, he applied the SIANN system primarily to candidate region classification and false positive reduction; in the commercial R2 system, he transformed it into a hybrid architecture combining full-image CNN scanning with a second-stage traditional algorithm filter, greatly improving efficiency. To adapt to the limited computing power of the 1990s, Zhang also introduced separable convolution kernels, splitting 2D convolutions into two 1D convolutions to drastically reduce computational load—a technique that later became common in efficient CNN design. R2's ImageChecker became the first FDA-approved radiology AI product. By the mid-2000s, it was assisting tens of millions of readings annually and driving an insurance payment market worth hundreds of millions of dollars. This article not only restores a forgotten piece of technology history but also reminds us of the true origins of fundamental AI innovations. For players and the industry, this…