In the race to create an AI that can be taught the same way as a human, China’s tech giant Baidu may have an edge: Its new virtual AI can learn skills the same way babies learn from their parents.
Baidu researchers taught an AI agent to navigate a 2D virtual world called XWORLD, using language and images to convey information to the agent. The virtual AI would then take that information to understand and complete the task at hand, and apply it to future skills using the language.
Baidu is one of the leaders in AI development, along with Google, Amazon, IBM, and others. As the “Google of China,” since the use of Google is still prohibited in China, it has a big edge: A massive trove of user data. In the fall of 2016, Baidu began offering an open-source platform for machine learning. It’s also been developing driverless technology in Beijing—which is heavily dependent on machine learning.
Baidu’s experiment aimed to tackle a major obstacle in AI development: Transferring skills from one domain to another, which is easy for humans, but difficult for AI. To overcome this hurdle, Baidu taught the AI agent language, in the same way humans would teach each other. The agent was given natural language commands—such as, “Please navigate to the apple,” or “Can you move to the grid between the apple and the banana?”—and was rewarded for completing the correct action. The result was that the AI agent could understand language that had first been meaningless to it, and could adjust actions accordingly.
According to Baidu’s release, the result “brings us one step closer to teaching machines to learn like humans do,” and the experiment could lead to development of a household robot that could learn from its “family.”
While the accomplishment is important in the development of an AI that can learn through language, it is not entirely novel. “After Andrew Ng announced his departure as the head of Baidu AI, Baidu is making sure people know they are still a top player in the field of AI,” said Roman Yampolskiy, head of the Cybersecurity Lab at the University of Louisville. In February 2017, Baidu shut down its two-year-old mobile health operation, merging it with its AI business and search business, in order to devote more resources to AI.
Baidu’s goal is to eventually teach a physical robot in a realistic environment using natural language. It plans to follow the current experiments with those using a virtual agent in a 3D environment, which is more realistic, and would pose new challenges to the AI. This work aligns with the company’s business goals: Its Baidu Duer is a virtual assistant that can search and mine information on the web via voice commands. Duer aims to extract meaning from information found on the internet by using an end-to-end, deep-learning system that can pick up English or Mandarin Chinese voices.