Suche
Mein Konto
AI study reveals: This is how the machine sees the human image!
Research from JLU Giessen and the Max Planck Institute analyzes AI perception of objects. Results published in Nature Machine Intelligence.
AI study reveals: This is how the machine sees the human image!
On June 23, 2025, a research team from Justus Liebig University Giessen and the Max Planck Institute for Cognitive and Brain Sciences published significant results on object recognition using artificial intelligence (AI). These findings were published in the renowned journal Nature Machine Intelligence published. The first authors Florian Mahner and Lukas Muttenthaler and the last author Prof. Dr. Martin Hebart presented a new approach to identifying and comparing key dimensions that both humans and AI pay attention to when seeing objects.
The study analyzed around 5 million odd-one-out judgments from 1,854 object images to find out which visual and semantic properties humans and AI prefer. It turns out that people focus on meaning-related dimensions, such as “animal-related” or “fire-related”, while AI models primarily focus on visual properties such as “round” or “white”. This phenomenon is called “visual preference” and could significantly impact trust in AI systems if there are differences in object recognition strategies.
Methodology and results of the research
The scientific work uses multiple deep neural networks (DNNs) to recognize images similar to humans and determine the key dimensions of the images. The comparison of dimensions between humans and DNNs revealed that although AI achieves approximations to these dimensions, it does not fully match human perception. What was particularly noticeable was that for animal-related dimensions, many non-animal images were not included in the analysis, which further influenced the results of the AI technology.
The researchers hope that future projects will enable a direct comparison between human and AI perception, which could lead to a better understanding of AI perception and improvement of the technology itself. Contact with Prof. Dr. Martin Hebart is mentioned here as a possibility for those interested in further information on this field of research.
Applications of AI in image recognition
The findings about the differences in object recognition have wide application in various areas. AI is often used to support logistics, image classification or customer structure analysis in e-commerce. In this context offers Teachable machines a tool for quick and easy programming of AI systems, which enables computers to recognize images, sounds or poses.
Example applications include support with inventory or sorting goods. The importance of the training data qualifies as crucial for the performance of the AI models. A practical example: If an Alexa is visible in the background of a training scene, the AI is tricked into classifying the image based solely on that presence, regardless of the primary focus of the image.
Future challenges and developments
AI technology, particularly in image recognition, shows enormous potential, but also has challenges to overcome. Concerns about data protection, bias in training data and the need for clear legal frameworks are key issues that need to be addressed. In addition, AI models must be robust to various conditions such as lighting and noise to work effectively in practice.
As developments in machine learning and neural networks continue, the future of AI-powered image recognition remains bright. Companies can benefit through process optimization and targeted marketing strategies, while extensive research into improving AI capabilities continues.