Reproduce “TIC-TAC: A Framework for Improved Covariance Estimation in Deep Heteroscedastic Regression” 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
Megh Shukla, Mathieu Salzmann, Alexandre Alahi
Researchers from EPFL introduced TIC-TAC, a framework to improve covariance estimation in deep heteroscedastic regression by proposing the Taylor Induced Covariance (TIC) for modeling prediction randomness and the Task Agnostic Correlations (TAC) metric for evaluation. This approach achieved superior correlation learning and improved negative log-likelihood convergence across synthetic, UCI, and 2D human pose estimation datasets.
Key insights
- Parameterizing covariance by relating it to the local behavior (gradients and curvature) of the mean function provides a more principled and interpretable way to capture prediction uncertainty.
- Explicitly modeling full covariance matrices, rather than just diagonal variances, is crucial for accurately capturing complex inter-variable correlations in multivariate regression tasks.
- Conditioning on partial observations offers an effective and task-agnostic approach to quantitatively evaluate the accuracy of learned correlation structures when explicit covariance ground truth is unavailable.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.