Google open-sources TensorFlow

Background

Training a machine learning model at scale has traditionally required a lot of machinery: many computers working together, code written specifically for that distributed setup, and people who can turn a research idea into a program that actually runs. Google built such a system for its own internal use, called DistBelief, and TensorFlow was described by Google as its second-generation machine learning system, built after that earlier experience^[1].

A machine learning framework is, in plain terms, a toolkit that packages the recurring steps of the work: describing a model, feeding data through it, measuring the error, and adjusting the model's parameters to reduce that error. The framework handles the bookkeeping and the mechanics of running the computation, so the person writing the model can concentrate on the model itself. Whether such a toolkit is available outside the company that built it makes a large difference to who can take part.

What happened

On 9 November 2015, Google Research announced that TensorFlow was being released as open source under the Apache 2.0 licence^[1][2]. The TensorFlow Blog gives the same date for the open-source release^[3], and the technical white paper accompanying the release is dated 9 November 2015 as well^[2]. Apache 2.0 is a permissive licence: it allows others to use, modify and redistribute the code, which is what makes a release usable by people outside the originating organisation rather than merely readable.

Google also stated why it was doing this. It said it hoped that open sourcing would let the machine learning community exchange ideas through working code rather than only through research papers^[4]. In other words, the aim was not just to publish a description of a system but to hand over the system itself, so that a method someone describes can be tried, compared and built upon by others directly.

Key details

Google's announcement stressed the range of hardware the same system could cover: TensorFlow could run on a single smartphone or across thousands of computers in data centres^[4]. That span matters, because models are often developed on large machines and then need to be served to users on much smaller ones.

Google also said it was already using TensorFlow in its own products at the time of the release: for speech recognition in the Google app, for Smart Reply in Inbox, and for search in Google Photos^[4]. The open-source release therefore came with working, shipped applications behind it rather than being a purely experimental code drop.

Impact

The most direct consequence of an open-source release is that the code can be inspected, reused and extended by anyone who accepts the licence, including students, researchers at other organisations and companies building their own products. Because the released code was the same system Google described as running in its own products^[4], outside users did not have to start from a description in a paper and reimplement it.

Google later published figures on how far the surrounding ecosystem had spread. A 2022 post on the TensorFlow Blog said that by then TFLite, the version of the system intended for smaller devices, ran on four billion devices, and that TensorFlow.js, the version that runs in a web browser, was being downloaded 170 thousand times weekly^[3]. These are Google's own figures about its own ecosystem, and they describe adoption rather than the technical merits of the system; but they do indicate that a toolkit released in 2015 ended up running in a very large number of places, from phones to browsers to data centres.