WHY THIS MATTERS IN BRIEF
As babies humans often “just learn” things, it’s a gift, but now AI’s are catching us up and starting to learn in the same way we do and it’ll change AI forever.
Given the glut of artificial intelligence (AI) breakthroughs in the past year, from AI’s building more AI’s, AI’s that can create and code their own programs, and AI’s that now learn as fast as humans do – and those are just for starters – it’s sometimes tough to pinpoint events that are especially pertinent. And this is one of those times.
Earlier this week, for example, I covered OpenAI’s “breakthrough” where their AI “just evolved” into an unsupervised learning system, and now Baidu have announced their own breakthrough where an AI agent that was taught English by a virtual teacher ended up being able to use natural language it had never seen before – and that’s the key here – to navigate its way around a video game.
Okay, so far kind of so boring. However, this means that Baidu, as per their announcement, have managed to create a system that’s managed to achieve fabled “Zero shot learning” which is the ability of an AI to solve a learning task without ever having been trained on it. And so far that type of learning capability, even for the world’s best AI researchers and teams, from DeepMind to OpenAI, haven’t been able to crack.
The Baidu experiment took place in an 2D maze like environment called XWORLD, a game where you have to kill a dragon to win, and their AI agent had to navigate its way around using just the natural language commands issued to it by the virtual English teacher.
In the beginning, the agent didn’t know anything about the language – every word was equally meaningless, because, for example, if you’re in a maze and someone says “Mind the dragon” that’s meaningless unless you know what a dragon is, where it is and what context the word “mind” is being used in.
However, as the agent explored the environment, the virtual teacher gave positive and negative feedback, then to help the agent learn faster, the teacher asked it some simple questions about the environment while it was navigating its way round. During the experiment the agent needed to correctly answer the teachers questions, then, by rewarding correct actions and penalising incorrect ones the teacher managed to train it to understand never seen before natural language and sentences, and as easily as that it “just learned” grammar – something that humans “just do” when we’re babies.
Zero shot learning. Voila, and while it sounds easy apparently it isn’t… but what do I know.
Humans use this form of learning all the time, and we’re great at taking a set of experiences from one situation and applying them to a never seen before situation, but now it looks like we might not be the only ones capable of learning in such a way and that’s a breakthrough. Score another one for AI.
