Meta-learning as a principle for human-like visual representations

2026-06-24

AI Summary

This work proposes a novel approach to learning more human-like visual representations by leveraging unsupervised meta-learning as a core mechanism. Unlike existing pretrained neural networks, which exhibit limited task flexibility compared to the human visual system, our method trains a sequential model on thousands of semantically rich image-to-concept mapping tasks. By integrating a distribution over higher-order semantic tasks, sequential modeling, and disentangled representation learning, the resulting representations significantly outperform baseline models. They achieve superior performance in predicting human similarity judgments, generalizing semantic rules, and aligning with neural activity in high-level visual cortex. These findings highlight the critical role of meta-learning in enhancing both task flexibility and alignment with biological vision systems.

Abstract

The structure of human visual representations underpins our capacity for adaptive behaviour. While pretrained neural networks model human visual representations with unprecedented success, a large discrepancy remains. We propose one reason: these networks optimise a single fixed objective, whereas human representations must support open-ended tasks. We hypothesise this flexibility arises from meta-learning (learning to learn), a pressure shaping representations to acquire new tasks from few observations. To test this, we train a sequence model, without any supervision from human data, across thousands of semantically rich tasks mapping images to high-level concepts. Compared to their pretrained base encoders, meta-learned representations better predict human similarity judgements, semantic rule learning, and high-level visual cortex. Behavioural gains depend on disentangled, high-level task distributions, while brain alignment is driven primarily by the learning-to-learn pressure. Our results suggest the flexibility of human visual representations reflects the functional demand to learn new semantic relationships on the fly.

Problem

• meta-learning
• visual representations
• human cognition
• task flexibility
• neural networks

Innovation

• meta-learning
• visual representations
• task flexibility
• neural alignment
• disentangled tasks