Alex Krizhevsky

In 2012, most computer vision researchers relied on traditional hand-crafted feature extraction methods. Alex Krizhevsky, then a PhD student at the University of Toronto under Geoffrey Hinton's supervision, bet on something almost nobody took seriously: training a deep neural network directly on raw images, using graphics cards (GPUs) as the computational engine. That bet, going against the grain of the entire field, turned out to be the most important moment in modern AI.

In October 2012, AlexNet (informally named in his honor) won the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) by a margin nobody expected: a top-5 error rate of 15.3%, compared to 26.2% for the second-place entry. That was a gap of nearly 11 percentage points, at a time when typical year-over-year improvement was just one or two. The paper he presented alongside Ilya Sutskever and Hinton at NeurIPS 2012, "ImageNet Classification with Deep Convolutional Neural Networks," now has over 120,000 citations on Google Scholar and stands as one of the most influential scientific works of the 21st century.

That same year, Krizhevsky, Sutskever, and Hinton incorporated DNNresearch Inc. The company was acquired by Google in March 2013, and Alex joined the Google Brain team as a research scientist, where he worked until September 2017. Before AlexNet, he also created the CIFAR-10 and CIFAR-100 datasets, two image collections still used universally today to train and benchmark computer vision models.

What I find fascinating about Krizhevsky is not just the outcome, but the clarity of the bet. While the field debated theory, he got his hands on the code, got hold of two GeForce GPUs, and simply tried. That lesson is more relevant than ever: sometimes the revolution doesn't come from an abstract idea, it comes from someone willing to execute what everyone else has been putting off. Every image an AI recognizes today, from your phone to self-driving cars, carries a bit of the stubborn work of this engineer who believed in the raw power of neural networks when very few did.