By using their newly developed algorithms, researchers are finally able to put any existing AI system to a test and also derive quantitative information about them: a whole spectrum starting from naive problem solving behavior, to cheating strategies up to highly elaborate "intelligent" strategic solutions is observed.
Dr. Wojciech Samek, group leader at Fraunhofer HHI said: "We were very surprised by the wide range of learned problem-solving strategies. Even modern AI systems have not always found a solution that appears meaningful from a human perspective, but sometimes used so-called 'Clever Hans Strategies'."
Clever Hans was a horse that could supposedly count and was considered a scientific sensation during the 1900s. As it was discovered later, Hans did not master math but in about 90 percent of the cases, he was able to derive the correct answer from the questioner's reaction.
The team around Klaus-Robert Müller and Wojciech Samek also discovered similar "Clever Hans" strategies in various AI systems. For example, an AI system that won several international image classification competitions a few years ago pursued a strategy that can be considered naïve from a human's point of view. It classified images mainly on the basis of context. Images were assigned to the category "ship" when there was a lot of water in the picture. Other images were classified as "train" if rails were present. Still other pictures were assigned the correct category by their copyright watermark. The real task, namely to detect the concepts of ships or trains, was therefore not solved by this AI system -- even if it indeed classified the majority of images correctly.
The researchers were also able to find these types of faulty problem-solving strategies in some of the state-of-the-art AI algorithms, the so-called deep neural networks -- algorithms that were so far considered immune against such lapses. These networks based their classification decision in part on artifacts that were created during the preparation of the images and have nothing to do with the actual image content.
"Such AI systems are not useful in practice. Their use in medical diagnostics or in safety-critical areas would even entail enormous dangers," said Klaus-Robert Müller. "It is quite conceivable that about half of the AI systems currently in use implicitly or explicitly rely on such 'Clever Hans' strategies. It's time to systematically check that, so that secure AI systems can be developed."
With their new technology, the researchers also identified AI systems that have unexpectedly learned "smart" strategies. Examples include systems that have learned to play the Atari games Breakout and Pinball. "Here the AI clearly understood the concept of the game and found an intelligent way to collect a lot of points in a targeted and low-risk manner. The system sometimes even intervenes in ways that a real player would not," said Wojciech Samek.
"Beyond understanding AI strategies, our work establishes the usability of explainable AI for iterative dataset design, namely for removing artefacts in a dataset which would cause an AI to learn flawed strategies, as well as helping to decide which unlabeled examples need to be annotated and added so that failures of an AI system can be reduced," said SUTD Assistant Professor Alexander Binder.
https://www.sciencedaily.com/releases/2019/03/190312103643.htm
