PaddlePaddle

What it is

PaddlePaddle (Chinese name 飞桨) is a deep learning framework that Baidu released as open source. A framework of this kind supplies the reusable parts of a neural network project: layer definitions, optimisation algorithms, data loading, and the plumbing that spreads training across many machines. Developers describe a model and the framework executes it, so a team does not have to rebuild that infrastructure for every project.

Where it came from

It began as an internal system. Its predecessor was Paddle — Parallel Distributed Deep Learning — a deep learning platform Baidu developed in-house from 2013 ^[2]. On 1 September 2016, Baidu chief scientist Andrew Ng announced at the Baidu World conference that the platform was being opened up under the name PaddlePaddle ^[2]. Baidu's own site dates the open-source release to September 2016 ^[1]. The public GitHub repository was created slightly earlier, on 15 August 2016, and is released under the Apache-2.0 license ^[3].

China News Service described the launch as making PaddlePaddle China's first open-source deep learning platform, and noted that Baidu thereby became the fourth company, after Google, Facebook and IBM, to open-source AI technology ^[2].

What the first release did

The launch version supported both CPU and GPU, in single-machine mode and in distributed mode, with training able to span hundreds of machines ^[2]. That distributed capability is the point of the original Parallel Distributed name: large models and large datasets do not fit on one server, so the computation has to be split up and the results recombined.

Rather than shipping only a bare framework, the first release included five ready-to-run tasks: neural machine translation, recommendation, image classification, sentiment analysis and semantic role labelling ^[2]. Each is a standard problem shape. Neural machine translation maps a sentence in one language to a sentence in another; recommendation predicts what a user is likely to click on or buy; image classification assigns a picture to a category; sentiment analysis judges whether a piece of text is positive or negative; semantic role labelling identifies who did what to whom inside a sentence. Offering these as runnable examples gave new users something that worked end to end, and pointed at the kinds of workloads Baidu itself cared about — search, advertising and content feeds all lean on ranking and on text understanding.

Where it sits

Releasing a framework openly is also a way of building a community around it. Frameworks are judged by the models people publish for them, the hardware they run on, and the number of people who can answer questions about them, so an internal platform that is never shared stays limited to the teams that built it. Baidu's decision to open its platform put it alongside the frameworks that Google, Facebook and IBM had already released ^[2], and gave Chinese developers a domestic option for the base layer of their machine learning work.

The project has grown well beyond its launch scope. Baidu's site now reports 4.77 million developers, 560,000 models created on the platform, and 180,000 enterprise and institutional users ^[1]. Those numbers describe not only a piece of software but an ecosystem: people training models with it, and organisations depending on it.