China’s Baidu Releases AI Code

This post is also available in: עברית (Hebrew)

Google and Facebook are not only some of the most sought out companies to work for, they also release just enough code and core technology to maintain their positions as industry leaders, around whose technologies all the rest coalesce. China’s Baidu, the country’s leading search engine, has not only succeeded in dominating the Chinese market, but is giving Google a run for its money in terms of technological development, with a particular focus on artificial intelligence (AI).

A fortnight ago, Baidu released some vital code for its deep learning technology, a powerful and increasingly popular machine-learning paradigm. Recently, Baidu used the code to develop an impressive speech-recognition technology called Deep Speech 2. In particular cases, specifically short sentences, Deep Speech outperforms most humans in recognising speech correctly. For Baidu, this is critical. Typing Chinese characters is particularly awkward on mobile platforms, and many in China already prefer using their voice to issue commands to the phone or send short messages.

Deep learning allows machines to transcribe voice, recognise objects in images, and a whole host of impressive feats. It works by simulating a large neural network that is consistently fed input. Over time, the neural networks “learns” to recognise patterns and almost any new example.

Baidu’s code, Warp-CTC, is an improved implementation of existing deep learning techniques. The code on its own, however, is not all that useful. Beyond the code, what’s needed to create a powerful AI is data; specifically user data that companies such as Google, Facebook, and Baidu possess in troves. Releasing the code, then, is not only not harmful to the companies, but even beneficial. First, they get to set the standard that everyone will work with. Almost as important is the research done on the code by members of the public, from which the companies benefit tremendously.