THERE IS A NEW PROGRAM THAT CAN TEACH AI COMMON SENSEDate: 2019-02-07
There is a new game from the Allen Insitute for Artificial Intelligence, which hopes to enlighten some of the algorithms of AI to teach these computers a better understanding of what common sense is.
Teaching computers how the world fits together can become a tricky business. No matter how much data you feed to an AIs algorithm, there are certain types of knowledge that just cannot easily be written down. Answers to questions like, "Who should I go for drinks with this evening: my friends, a lion, or the President?" and "Can I eat a piece of steak with a straw?" (The answer to these questions might be "Yes, but only with a lot of patience.")
However, AI researchers say they have an answer to this, which comes in the form of a tool that might assist in teaching AI common sense. That tool is Pictionary. Yes, Pictionary! Scientists from the Allen Institute for Artificial Intelligence (AI2) released an online Pictionary-style game that goes by the name of Iconary, which they say could help sharpen AI's common sense.
In Iconary, players must illustrate complex scenes like "a crowd celebrating a victory" or "kicking a tennis ball on a beach" while a computer program named 'AllenAI' tries to guess what they've drawn. It is importantly to note that the AI system has never seen these phrases before, but for humans it is common sense.
The reason this is challenging for AI, says AI2 researcher, Ani Kembhavi, is because it tests a wide range of common sense skills. The algorithms must first identify the visual elements which is given in the form of a picture, then the AI should figure out how it relates to these drawings, and then translate that scene into simple computer language that the AI can understand. This is why Pictionary could teach computers information that other AI systems, like Go and StarCraft, can't.
"One of the things about [board games and video games] is that they’re a little bit removed from reality," says Kembhavi. "The breakthroughs they help obtain are phenomenal, but they're not directly applicable to real life. Whereas if you want to do well at a game like Pictionary, you need to have common sense knowledge and you need to be able to reason with that knowledge."
Take the phrase "friends chatting in a kitchen," for example. To depict or guess this phrase, you not only need to understand the elements that makes up what friends are, and those that comprise a kitchen (pots, pans, a fridge, some food), but you also need to know how to show that the friends are chatting and not eating. So you might want to avoid placing them around a table and so on. It's a lot of information to take in.
This is why a system like Iconary is different to another AI Pictionary game that Google released back in 2017 named Quick, Draw! While the latter is focused primarily on recognising individual doodles, AI2's game is all about combining these elements into something more complex.
Iconary is not going to give computers the gift of common sense overnight, obviously, but AI2 hopes the data it collects will be useful for all the future experiments to come.
Collecting data from humans is particularly important for Pictionary as AI systems can't play the game against one another to improve. With games like StarCraft, researchers speed up the training process by challenging AI against AI in a sped-up environment. Each agent learns individual strategies within the confines of the game, and the researchers pick the most successful agent to go forward with.
Because Iconary is trying to extract as much data as possible, which is fundamentally unknown to machines, it still needs humans in the loop. "If you got AllenAI to play itself, it would develop a communication protocol that only it knows," says Kembhavi. Basically to translate it into English, to speed up the system's guessing process, it would learn, over time, to sub in shorthand symbols for complex phrases. So this symbol '҂' might mean 'drinking orange juice," for example. "And then you'd deploy it in the real world, and you wouldn't understand what it means," says Kembhavi.
This might be a comforting thought for those who are still worried about AI systems taking over the human race. Not only do computers not yet understand the world as we do, but they need our help to improve. You can view the Iconary game in action in the video below.