Mensch besser als KI: KI hat gravierende Schwächen beim Erkennen von Formen
Warum KI selbst die einfachsten Umrisse oft nicht erkennt – und Katzen plötzlich als „Kreuzworträtsel“ deutet.
Recent research has revealed that humans possess a significant advantage over artificial intelligence (AI) when it comes to recognizing shapes and forms. Even when objects are presented as simple silhouettes or shadows, humans can still identify them accurately, while AI faces considerable weaknesses in this area, according to a study conducted by researchers from New York University Grossman School of Medicine.
The journal "iScience" reports that the underlying reason for this gap lies in the distinct ways in which the human brain and modern deep neural networks process visual information. While humans prioritize overall shape, AI tends to focus on local details, surface patterns, and textures. Even minor distortions in images or the absence of typical details quickly cause the AI systems to fail.
Over 200 different types of deep neural networks were tested, all of which fell short of human performance across various conditions. In one particular experiment, researchers obscured the outlines of objects and animals with numerous small crosses. While humans could still identify the objects based on the outer contours, most AI models failed completely, misidentifying cat or butterfly silhouettes as crossword puzzles, window grates, turtleneck sweaters, or tent worms.
Current AI models, despite being trained on vast amounts of photos with human labeling, do not function as closely to human vision as previously thought, even when combined with textual descriptions. Although more advanced AI systems that incorporate both images and associated text tend to perform better overall, they lose their edge when the overall shape is disrupted or evaluated in isolation.
The significance of these findings for practical applications is substantial. To ensure reliable functioning of AI systems in robotics, autonomous driving, and vision prostheses or brain-computer interfaces for the visually impaired, especially in challenging conditions like fog, low light, or occlusions, the research suggests that future algorithms must be trained to develop a more holistic and robust visual capacity.
Written by urgent.news from Handelsblatt's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
