Clarifai

Contribution to AI

A year after AlexNet, the open question was whether that result had been a fluke of one architecture and one team. The 2013 ImageNet classification challenge answered it: the winning entry, submitted under the Clarifai name, cut top-five error to 11.7 per cent, and 11.2 per cent using additional training data. More useful than the score was the method behind it. Zeiler and Fergus had built a deconvolutional technique for projecting the activations of a hidden layer back into pixel space, so one could see which patterns a given feature map responded to. That diagnosis pointed directly at the changes — smaller first-layer filters, a shorter stride — that produced the winning network, and it turned architecture design from guesswork into something inspectable. The feature-visualisation strand of interpretability research starts there. The commercial move followed in October 2014, when the model was opened to developers as a metered API: send an image, receive ranked concept tags, with no GPU, no training data and no machine-learning background required. That preceded the equivalent services from Microsoft, Google and Amazon by one to two years, and established the pattern they copied — recognition sold per call rather than per licence, later extended to custom models trained on a customer's own small set of labelled images. Its Project Maven contract also made a startup, rather than a giant, the test case for military computer vision.

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