Reproduce “Transformers as Meta-Learners for Implicit Neural 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
Yinbo Chen, Xiaolong Wang
Researchers at UC San Diego developed a Transformer-based hypernetwork that directly infers the complete weights for an Implicit Neural Representation (INR) from observations in a single feed-forward pass. This approach enables high-fidelity 2D image regression (e.g., 31.96 PSNR on CelebA) and 3D view synthesis, surpassing gradient-based meta-learning baselines without requiring test-time gradient optimization.
Key insights
- Transformers can directly infer the complete set of weights for an INR, effectively bypassing the single-vector information bottleneck common in hypernetworks and eliminating test-time gradient optimization.
- The iterative nature of Transformer's residual updates can be conceptually linked to gradient-based meta-learning, offering a generalized view of these learning paradigms.
- The generated INRs implicitly exploit underlying data structures, with different weight columns attending to distinct, structured parts of the input data, suggesting a deeper understanding of input regularities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.