Researchers developed a new way to define visual similarity using Anchor Divergence
A new method allows AI to understand that similarity between images depends on the specific context, such as style or clinical details.
Current AI models often struggle because they see similarity in a fixed way. When an AI compares two images, it usually relies on a single geometric measurement. This often ignores that what makes two pictures similar can change depending on your goal.
A new paper from October 2026 introduces a concept called Anchor Divergence to solve this. Instead of a one-size-fits-all approach, this method uses probability distributions to adapt how the AI measures similarity.
Contrastive representations naturally encompass a family of geometries that can be specialized to particular semantic structure.
Imagine you are looking at a database of medical scans. Sometimes you want to find images that look physically similar, but other times you want to find scans that show the same type of injury. The context of your search matters.
By using Anchor Divergence, researchers can specify these different contexts. The system then adjusts its internal geometry to prioritize the features that matter most to the current task.
This technique links contrastive learning, where models learn by comparing pairs of items, with information geometry. It allows the model to treat similarity as a flexible tool rather than a rigid rule.
Think of it like choosing different lenses for a camera to focus on specific details, rather than just zooming in and out. How might this precision change the way you organize your own digital photo library?
It is important to note that this method was tested specifically on retrieval tasks. The researchers have not yet applied it to generative models or other broader areas of machine learning. The effectiveness of this approach relies on having well-defined anchor distributions for the specific context required.
Source: arXiv cs.AI