The Chinese Internet giant Baidu Inc. has been makingbig progressinapplying deep learning neural networksto improve image recognition, language translation, search ranking and click prediction in advertising. Now, it’s going to give a lot of it away.
The company, often called “China’s Google,” will announce Thursday at the annual Baidu World conference that it’s offering the artificial intelligence software that its own engineers have been using for years as open source onGitHub. It’s code-namedPaddlePaddle, for PArallel Distributed Deep LEarning.
Deep learningis the branch of machine learning that attempts to emulate the way neurons work in the human brain to find patterns in data representing sounds, images, and other data. Google, Facebook, Microsoft, IBM and other companies have also been making big breakthroughs thanks to the ability to pump massive amounts of data into these artificial neural networks.
The announcement follows the open-sourcing in the last two years of other machine intelligence and deep learning tools such asTorchandmachine-vision technologyfrom Facebook,TensorFlowfrom Google andComputation Network Tool Kit(CNTK) from Microsoft, as well as independent open source frameworks such asCaffe. Baidu also has open-sourcedother pieces of its AI code. But Xu Wei, the Baidu distinguished scientist who led PaddlePaddle’s development, said this software is intended for broader use even by programmers who aren’t experts in deep learning, which involves painstaking training of software models.
“You don’t need to be an expert to quickly apply this to your project,” Xu said in an interview. “You don’t worry about writing math formulas or how to handle data tasks.” (Indeed, the playful doubling of the original code-name is intended to convey that it’s easier to use than rival software.)
It clearly requires acertain base of knowledge, but Xu said PaddlePaddle requires significantly less code than some alternatives. For instance, a machine translation model built on it needs about a quarter of the specially written code than other AI platforms require, he claimed.