Form versus Texture
Modern image recognition programs usually work impressively accurately. But sometimes they make embarrassing mistakes. This could be because the algorithms focus on the texture of objects instead of their shape.
One of the greatest strengths of current artificial intelligences (AIs) is the classification of images - some of them even surpass humans in some respects. However, the slightest change can occasionally throw these algorithms off course. For example, if you edit images so that a cat's fur appears tabby, spotted, or spotted, humans still have no problem identifying the animals in most cases, while machines fail completely.
Researchers from the University of Tübingen have now discovered the possible reason for this, as they explained at the "International Conference on Learning Representations" (ICLR) that took place in May 2019: While humans mostly pay attention to the shape of an object, machines focus learning algorithms on their texture. This surprising result shows how differently humans and machines "perceive" things…