Deterministic convolutional neural networks (CNNs) that guarantee equivariance to rotations and translations — a property known as SE(3)-equivariance — have become a powerful tool in 3D computer vision, but they lack the ability to quantify uncertainty in their predictions. That limitation is a critical barrier for adoption in safety-critical applications such as autonomous navigation, medical image analysis, and robotics, where knowing how confident a model is about its output can be as important as the output itself.
A new paper from researchers including Keripale, Abhishek, Thiagarajan, Ponkrshnan, Ghosh, and Susanta proposes a Bayesian Steerable-CNN that bridges this gap. The work, available on arXiv, introduces a method to place posterior distributions over the basis coefficients of steerable kernels, yielding stochastic kernels while preserving exact SE(3)-equivariance.
The Problem: Uncertainty in Equivariant Networks
Steerable CNNs achieve equivariance by parameterizing convolutional kernels as linear combinations of steerable basis functions. This ensures that rotating or translating the input produces a corresponding transformation of the feature maps. However, typical steerable CNNs are deterministic: they output a single point estimate with no confidence measure. According to the paper, this "precludes uncertainty quantification — limiting their use in settings where confidence estimates are essential."
The authors address this by adopting a Bayesian approach. Instead of learning fixed kernel coefficients, they learn posterior distributions over those coefficients. The loss function is derived via variational inference and minimized using Bayes-by-Backpropagation. The framework further decomposes predictive uncertainty into two components: epistemic uncertainty (due to model uncertainty) and aleatoric uncertainty (inherent noise in the data).
Key Results: Accuracy and Calibration
The Bayesian Steerable-CNN achieves competitive classification accuracy alongside strong calibration. The model attains an expected calibration error (ECE) of 0.0263, indicating that its confidence estimates closely match true accuracy. Under distributional shift — simulated by additive Gaussian noise — the Bayesian version outperforms its deterministic counterpart by up to 6.17%.
Moreover, the paper reports that by leveraging the uncertainty estimates, the model improves overall accuracy: it achieves roughly 4% higher accuracy across 84% of the test dataset. A statistically significant negative correlation between epistemic uncertainty and prediction error confirms that the learned posterior variance is semantically meaningful — when the model is uncertain, it is more likely to be wrong, and vice versa.
| Metric | Value |
|---|---|
| Expected Calibration Error (ECE) | 0.0263 |
| Accuracy gain over deterministic (under noise) | up to 6.17% |
| Accuracy gain on 84% of test set | ~4% |
Implications for Industry
While the paper is a pure research contribution without direct industrial application, the underlying technology has clear relevance for enterprise systems that rely on 3D perception and require both geometric consistency and reliable confidence estimates. For example, in autonomous logistics, a robot picking items from a bin must not only recognize objects correctly regardless of orientation (equivariance) but also know when it is unsure (uncertainty). Bayesian Steerable-CNNs offer a principled way to achieve both without sacrificing the inductive bias that makes equivariant models sample-efficient.
The fact that the model maintains exact equivariance while adding posterior distributions over kernel coefficients means that existing architectures can potentially be upgraded with Bayesian uncertainty quantification without changing their core equivariance guarantees. The use of standard variational inference techniques (Bayes-by-Backpropagation) also suggests that the approach can be implemented with widely available deep learning frameworks.
Future work may extend this framework to other domains such as point cloud processing, 3D medical imaging, or even 2D applications where equivariance to rotations is important. The paper's demonstration that epistemic uncertainty correlates meaningfully with error indicates that the model not only quantifies uncertainty but does so in a way that can be acted upon — for example, by rejecting low-confidence predictions or triggering human review.
"The framework unifies Bayesian uncertainty quantification with the inductive bias of equivariant CNNs."
The authors have made the paper available under a Creative Commons license, and the code and data are expected to be released. As of the paper's publication date of June 13, 2026, the work represents a step toward building deep learning systems that are both geometrically aware and trustworthy.