Reproduce “Meta-learning as a principle for human-like visual representations” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Can Demircan, Marcel Binz, Alireza Modirshanechi, Eric Schulz
Meta-learning, when applied to a causal Transformer learning diverse tasks from Sparse Autoencoders, yields visual representations that enhance alignment with human similarity judgments, sequential learning behaviors, and fMRI brain activity in high-level visual regions. The work demonstrates that training for rapid adaptation, rather than fixed objectives, can explain the structure of human-like visual cognition.
Key insights
- The functional demand for rapid adaptation, formalized as meta-learning, intrinsically leads to visual representations that align more closely with human cognitive and neural organization without explicit human data supervision.
- The 'learning-to-learn' pressure from meta-learning, distinct from mere exposure to diverse tasks (multitask learning), is crucial for achieving human-like representations, particularly for brain activity alignment.
- Meta-training on tasks defined by disentangled, high-level concepts from SAEs is more effective at inducing human-like behavioral alignment than using entangled or mid-level features.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.