University of Toronto

Contribution to AI

Funding is the part of this story that usually goes unrecorded. Hinton left Carnegie Mellon in 1987 for a CIFAR fellowship in Toronto, and what the money bought was continuity — a small group free to keep working on backpropagation and learned representations through the years when the approach was considered a dead end and grant committees elsewhere said so plainly. Yann LeCun spent a postdoctoral year there before going to Bell Labs; the pattern of arriving, absorbing methods nobody else taught, and leaving to seed them repeated for three decades. The technical payoff arrived in stages. Two 2006 papers showed that deep networks could be trained layer by layer after all, which made depth a practical question again. Between 2009 and 2012 Toronto students replaced the Gaussian mixture models that had dominated acoustic modelling for a generation, and the improvement was large enough that speech groups at Microsoft, Google and IBM converged on the same conclusion. Then AlexNet took the 2012 ImageNet contest with a top-five error of 15.3 per cent against 26.2 for the next entry, trained on two gaming GPUs, and hand-designed vision features stopped being worth writing. Google's purchase of the three-person DNNresearch in March 2013 set the price and the template for every academic acqui-hire that followed, and the Vector Institute of 2017 turned the accident of one funded fellowship into permanent national infrastructure.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).