Reproduce “Prediction under Latent Subgroup Shifts with High-Dimensional Observations” 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
William I. Walker, Arthur Gretton, Maneesh Sahani
The partial RPM adapts image prediction when hidden subgroup prevalence changes between training and deployment. It learns subgroup information from source-linked concepts and proxies using recognition networks, then estimates the target mix from unlabeled images without reconstructing them. In constructed experiments, accuracy improves with more CIFAR-10-based data and withstands added noise, but this does not establish naturally occurring shifts.
Key insights
- Avoiding explicit generative modeling and reconstruction of high-dimensional observations (like images or noisy features) is crucial for scaling latent shift adaptation methods to complex real-world data.
- By focusing solely on learning effective recognition mappings from high-dimensional observations to a low-dimensional latent space, the model becomes robust to irrelevant or noisy dimensions in the input data.
- The Recognition-Parametrised Model provides a suitable framework for identifying and adapting to unobserved confounding latent variables, especially when these latents are discrete, through exact maximum-likelihood learning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.