DNNresearch Inc.

What it was

DNNresearch Inc. was a company incorporated in 2012 by Geoffrey Hinton, a professor at the University of Toronto, together with two of his graduate students, Alex Krizhevsky and Ilya Sutskever^[1]. The University of Toronto described it as a start-up from its Department of Computer Science^[1]. Its field was deep neural networks: machine-learning systems built from many stacked layers of simple units, where each layer transforms what the layer below it produced. The company's research covered speech recognition, computer vision and language understanding^[1].

Where it came from

The company grew out of work done in Hinton's research group. The university said that Krizhevsky and Sutskever developed a system that dramatically improved the state of the art in object recognition^[1] — the task of having a program say which object appears in an image. Systems like this are trained rather than written by hand: instead of coding rules for what a chair or a cat looks like, researchers show the network many labelled examples and adjust its internal parameters so that its answers move closer to the correct ones. Improvements of the kind the university described usually come from a better architecture, more data, more computing power, or some combination of the three.

The group's interests were broader than images. The company's research also covered speech recognition and language understanding^[1].

The acquisition

The University of Toronto said the company had been acquired by Google for its research on deep neural networks^[1]. Financial terms of the deal were not disclosed^[1].

The purchase did not come out of nowhere. Google had earlier given Hinton's research group a $600,000 gift through its Focused Research Awards^[1], a company programme that gives money to academic research groups. In other words, the two sides already knew each other's work before any acquisition was discussed.

People rather than products

iThome reported that the company's website showed nothing but its name, and did not describe its research or results^[2]. That fits a company whose value lay in a handful of researchers and in what they might do next, not in a product already on sale.

What Google got, in practice, was the people. The university said Krizhevsky and Sutskever would both move to Google after the acquisition^[1], while the arrangement let Hinton divide his time between the university and Google^[1]. The professor kept his academic post and also worked with the company — a pattern that is not unusual in machine-learning research, where the same small number of people are sought after by both universities and industry labs.

Why it is remembered

DNNresearch sold nothing and said little in public. What makes it notable is the shape of the deal: a research group whose work on object recognition, speech and language was strong enough that a large technology company bought the group itself and kept its members working on the same problems inside the company. For anyone following how deep learning moved from university labs into industry, this is a small, concrete example of that movement.