Researchers at the Allen Institute, based in Seattle, USA, have developed a new version of the Iconary game capable of improving communication skills and establishing connections between different types of objects, in integrated artificial intelligence (AI) systems.
The machine learning algorithms used to create this more advanced variant of the game allow trained bots to play against each other or with humans. In addition, they are more adept at dealing with complex semantic aspects, associating words with corresponding figures.
“Our work is based on a system that aims to train models to play the game Iconary — a game based on Pictionary that we created in 2019 — in which one player has to guess what another is drawing. This new version uses modern AI methods to improve communication between artificial agents and humans”, explains engineer Christopher Clark, co-author of the project.
Want to stay on top of the best tech news of the day? Access and subscribe to our new youtube channel, Kenyannews News. Every day a summary of the main news from the tech world for you!
Iconary 2.0
Instead of building a game in which players compete with each other, scientists have developed a technique to improve cooperation and improve the visual communication of artificial intelligence devices. Although this semantic processing is an innate ability of human beings, neural networks have difficulty dealing with the interpretation of drawings and images.
In this new version of Iconary, the researchers trained the AI algorithms using more than 55,000 disputes between human players. Later, when testing the performance of these algorithms, they found that flesh-and-blood players still outperformed AI agents, especially with regard to the ability to convey objects or ideas through drawings.
“We tested the model with sentences that contained words it had not seen in human-to-human games. This means that the models could not play just by reusing the strategies observed during training. Overall, we were able to show that we can train AIs reasonably good at understanding the human-authored drawings,” adds Clark.
Interpret better than draw
The researchers believe that, in the future, Iconary could be useful in training AI algorithms, allowing to accurately assess the ability of semantic connection between texts and figures. By now, it’s become clear that virtual gamers are better at guessing implicitly communicated concepts than conveying them through drawings.
According to scientists, the game can also be used as a valuable tool to create more efficient artificial intelligence systems, capable of understanding and interpreting visual forms of communication present in people’s daily lives, such as street signs, icons or emojis.
“This shows that AIs can apply world knowledge to the task of drawing and understanding the figure at a deeper level than just memorizing the strategies they were trained with. They are still worse draftsmen than humans, but that’s just a matter of time,” predicts Christopher Clark.
