Meta-learning yields AI visual representations that mirror human perception

A new study from researchers Can Demircan, Marcel Binz, Alireza Modirshanechi, and Eric Schulz proposes that meta-learning—a process where models learn how to learn from few examples—is the key principle behind human-like visual representations. Published on arXiv (2606.28399), the work argues that while pretrained neural networks can model human vision with great success, they still lag behind because they optimize a single fixed objective. Humans, by contrast, must support open-ended tasks. To test their hypothesis, the team trained a sequence model on thousands of semantically rich tasks that require mapping images to high-level concepts—all without any human data for supervision.

The results are striking: meta-learned representations outperformed their pretrained base encoders in predicting human similarity judgments, semantic rule learning, and even activity in the high-level visual cortex. The researchers found that the behavioral gains depended on the use of disentangled, high-level task distributions during training, but the brain alignment was driven almost entirely by the meta-learning pressure. This suggests that the flexibility of human visual representations is a functional adaptation to the demand of learning new semantic relationships on the fly. The study bridges machine learning and cognitive science, offering a new lens for building AI that sees the world more like we do.