ANGLE framework beats traditional regression on circular data tasks

Circular data — angles or directions — is common in computer vision (pose estimation), biology (animal migration), and meteorology (wind direction). Traditional regression methods fall short because they target the conditional mean, which is geometrically misleading for multimodal or skewed circular responses. To solve this, a team of researchers (Pathak, Roy, Chakraborty) developed ANGLE (Angular Neural Generative Learning via Engression), a deep generative framework that learns the full conditional distribution of an angular response given Euclidean and circular covariates.

ANGLE is optimized via a novel generalized circular energy score (GCES) loss, with proven strict propriety and rotational equivariance. The framework supports both pre- and post-additive noise models, and provides a unified toolbox for previously underexplored challenges in circular statistics: extrapolation, sufficient dimension reduction, and conditional distribution equality testing. In benchmarks, ANGLE significantly outperforms traditional methods on object pose estimation (critical for surveillance and autonomous vehicles) and wind direction prediction (vital for energy systems). The model achieves superior predictive performance and robust uncertainty quantification, all while being computationally lightweight.