The AI research team of Baidu has once again reached a significant milestone. The folks at the artificial intelligence team were able to train a virtual agent who resides in a 2D environment. The said subject was taught how to move about using natural language commands. The most interesting thing about it is the agent's ability to develop what they call "zero shot learning ability", which basically means having a basic sense of grammar. All these were achieved by using human-like training which basically starts by teaching the subject language through both positive and negative reinforcement.
As per Tech Crunch, Baidu's manner of teaching the agent is very much similar to how humans teach their babies basic language and grammar. Technically, kids are shown images and repeatedly told different words. Eventually, with enough reinforcement, kids are able associate words with images.
Nevertheless, unlike humans, robots and computers are usually unable to apply its vast past knowledge in performing new tasks. For example, a computer or a robot can know what a dragon fruit looks like. However, it may not be able to cut the fruit using a knife within command unless it has been explicitly trained to do so. This is where Baidu's AI differs from other machines. It is able to use past knowledge to perform new tasks.
According to Baidu's team, their agent was able to transfer what they know about the appearance of a dragon fruit in completing the task of "cutting X with a knife". The team said that their subject successfully performed the task without the explicit training that most computers require before mastery. With that said, Baidu was able to show that systems can learn and apply knowledge in a way similar to how humans do it, even if their subject was still limited to the realm of simple and video game-like environment.
