New research suggests that the key to making artificial intelligence see like humans may be meta-learning — the ability to learn how to learn. Scientists at the Max Planck Institute for Biological Cybernetics trained a sequence model on thousands of tasks requiring it to map images to high-level concepts, without any direct human data. The resulting representations better predicted human similarity judgments and brain activity in the visual cortex than those from standard pretrained networks.
The Research
Can Demircan, Marcel Binz, Alireza Modirshanechi, and Eric Schulz from the Max Planck Institute for Biological Cybernetics published their study on arXiv on June 24, 2026. They tested whether meta-learning — a pressure to acquire new tasks from few observations — could explain the flexibility of human vision. The team trained a sequence model on thousands of semantically rich tasks, each linking images to high-level concepts. Compared to pretrained base encoders, the meta-learned representations improved prediction of human similarity judgments by 15–20% and better matched neural activity in the high-level visual cortex (areas like the fusiform face area). The behavioral gains depended on disentangled, high-level task distributions, while brain alignment was driven primarily by the learning-to-learn pressure itself.
Why It Matters
Human visual representations support open-ended learning: we can recognize a new object after seeing it once or grasp a novel concept from a few examples. Standard AI models, optimized for a single fixed objective, fail at this. This study suggests that meta-learning — learning to learn — may be the missing principle that makes visual representations both flexible and human-like. For your own cognition, it implies that training your brain on diverse, high-level tasks (such as learning new categories or rules) could strengthen your ability to adapt to new visual challenges.
What You Can Do
- Challenge your visual system with diverse tasks: try identifying objects in different contexts, learn new visual patterns, or practice categorizing unfamiliar images.
- Engage in activities that require learning new semantic relationships, such as puzzles, visual memory games, or strategy games that involve recognizing patterns.
Source: arXiv q-bio.NC
Curious about your own brain? Take our free adaptive IQ test or try 306 brain training levels.
