Bayesian 3D Steerable CNNs: Enabling Equivariance and Uncertainty Quantification Simultaneously
Researchers have developed a novel Bayesian 3D Steerable CNN that simultaneously achieves SE(3)-equivariance and quantifies uncertainty. This new model places posterior distributions over kernel coefficients, enabling stochastic kernels while maintaining exact equivariance. The framework decomposes predictive uncertainty into epistemic and aleatoric components, demonstrating competitive classification accuracy and improved performance under distributional shifts. AI
IMPACT Introduces a new method for uncertainty quantification in equivariant neural networks, potentially improving reliability in applications sensitive to confidence estimates.