Share
New Collaborative AI Game Sees Machine Learning Cooperate With Humans Rather Than Crush Them

New Collaborative AI Game Sees Machine Learning Cooperate With Humans Rather Than Crush Them

Until now, the aim of AI algorithms developed to play games has been to crush human opponents. Well known examples include Alphabet AI unit’s DeepMind’s AlphaGo and IBM’s Deep Blue. The former has vanquished the best human players of the board game Go, while the latter is the undisputed reigning chess champion. A new game with an AI player will go live tomorrow. But this time the machine intelligence-player’s role is one of collaborative team work in unison with a human player rather than being pitted against them.

The game is called Iconary and shares similarities with Pictionary. It has been developed by the Allen Institute for Artificial Intelligence, co-founded by the late Microsoft co-founder Paul Allen. Its aim is to provide a testing ground for how machine learning AI and humans can collaborate and reach mutual understanding on a level that improves that which either party would be able to achieve alone. Human and machine players have to learn to interpret each other.

Oren Etzioni head of the Allen Institute for AI explained the Iconary concept as:

“A lot of the [AI] research work hasn’t really explored collaboration. How do you build collaborative systems? We see this as a test bed for debugging interaction.”

The big weakness of current AI algorithms is their lack of ‘common sense’. This means that they have to be programmed for a specific task and are not able to understand when rules that apply to one set of circumstances would logically also apply to another. This lack of abstract common sense knowledge applicable across various scenarios severely limits the flexibility of AIs.

The Iconary AI has been ‘trained’, or given data input, on 100,000 drawing games played between human players. It recognises 1200 different ‘icons’. These are randomly combined into phrases such as ‘sneezing into soup’ or ‘woman lifting a table’. One side has to draw the phrase and the other to guess it. This exposes the AI to a huge number of possible ‘clues’ thought up by the human player that it has no previous experience of. As it encounters the kind of clues that the human mind creates as associated with a picture of a phrase consisting of several elements, the AI learns to better interpret its human partner.

It is hoped that the collaborative approach will lead to an AI that is able to much more intuitively interpret the sentiment of human requests or statements that don’t fall into its initial programming or experience. This would be particularly useful for voice-operated assistants such as those becoming popular at home like the Google Assistant and Amazon’s Alexa.

Iconary will be made freely available online in the hope of attracting large numbers of human players that the game’s AI will be able to learn how to interpret and guess the sentiment of. The ultimate goal is that the AI becomes able to display a level of ‘common sense’ when interpreting human desire and sentiment.

Leave a Comment