Teaching computers common sense is a tricky business, as researchers acknowledge that no matter how much information you feed to an algorithm, there is something incredible about human intelligence which simply cannot be written down. Scientists from Allen Institute of Artificial Intelligence wants to solve this AI problem with the help of a new tool which might help sharpen common sense of AI.
Allen Institute of Artificial Intelligence released an online game ‘Iconary’ styled after Pictionary. In Iconary game, players have to illustrate complex visual scenes by drawing as the AllenAI tries to understand the meaning of the scene and converts it to verbal sentences. The scenes could be like ‘’a crowd celebrating in a birthday party’’ or ‘’ kicking football on ground’’. The most important challenge to the AI is that it has never encountered such phrases before.
The AI starts by identifying visual elements in a picture and figure out relationship of individual parts with others in the picture and later translating that scene into verbal language understandable to humans. This capability of inculcating a broad range of common sense skills in the AI sets Iconary apart from other AI benchmarks like StarCraft and Go. The researchers were trying to overcome the breakthroughs obtained from video games and board games which are phenomenal but cannot be directly applied to real-life scenarios. Whereas a game like Iconary has the potential to derive common sense knowledge from the given data.
If we take a phrase ‘’ friends playing cricket on ground’’. To depict the scene, AI will have to understand the meaning of ‘friends’ and what humans conceive from the word, along with cricket ( cricket bat, ball, playground) and the AI will have to show the meaning of ‘playing’ through visual image. The complexity of the visual situation further muddles the AI’s understanding of it. Other versions of the Pictionary game such as Quick, Draw! released in 2017 were much simpler as they only focused on recognizing simple doodles while Allen’s Iconary game combines different elements and tries to derive a coherent meaning out of seemingly complex situations.
Aniruddha Kembhavi, a researcher at AI2 said that Iconary will extract data which is fundamentally yet unavailable to machines meaning that AI training needs humans to be in the loop. This seems to be a comforting news for those who are worried about AI making humans obsolete. Computers don’t understand themselves yet neither the human world and for that to happen they need human help.