Mobile-Menu

Study Small AI Models Recognize Images More Robustly

Source: Press Release from Osnabrück University | Translated by AI 1 min Reading Time

Related Vendors

Taiwan Machine Tool & Accessory Builders' Association (TMBA)
SIMTOS (Seoul International Machine Tool Show)

A research team from the University of Osnabrück shows that smaller AI models can be more robust in image recognition than large systems. The key factor is not the model size but the type of training.

AI models often process visual information in a fragmented manner – new training approaches address this.(Source:)
AI models often process visual information in a fragmented manner – new training approaches address this.
(Source:)

The study, published in *Nature Machine Intelligence*, introduces a biologically inspired training approach. The researchers draw inspiration from the development of human vision. Instead of immediately training with high-resolution images, the models initially go through a phase with reduced image quality, such as blurred or low-contrast representations. They refer to this training concept as a "Developmental Visual Diet."

The background is a well-known problem: artificial vision systems often recognize objects reliably but react sensitively to disturbances such as image noise, altered textures, or targeted manipulations. While humans focus more on shapes, many AI models primarily rely on surface patterns.

Robustness through Development-Oriented Training

The models trained in this way demonstrated higher robustness in tests. They recognized objects more reliably even with poor image quality and were more resistant to disturbances. Additionally, they based their decisions more on the shape of objects rather than textures. In individual tests, robustness increased significantly.

As an example, the research team describes an image that forms a bear motif from individual bottles: "The stimulus shown here is an example where traditional AI decides differently from humans. AI recognizes bottles, people see a bear," explains Zejin Lu, the study's lead author.

"Our results show that robust AI is not just a matter of more data, larger models, and higher computing power," says Tim Kietzmann from the University of Osnabrück (Germany). "What is also crucial is how a model learns. When artificial vision systems undergo a visual development more closely aligned with human learning, they also approach human vision."

Subscribe to the newsletter now

Don't Miss out on Our Best Content

By clicking on „Subscribe to Newsletter“ I agree to the processing and use of my data according to the consent form (please expand for details) and accept the Terms of Use. For more information, please see our Privacy Policy. The consent declaration relates, among other things, to the sending of editorial newsletters by email and to data matching for marketing purposes with selected advertising partners (e.g., LinkedIn, Google, Meta)

Unfold for details of your consent

Related content