Bayesian Topological Convolutional Neural Nets

Sarah Harkins Dayton, Hayden Everett, Ioannis D. Schizas, David L. Boothe, Vasileios MaroulasarXiv 2025 · DOI · arXiv

Abstract

Standard Convolutional Neural Networks (CNNs) often struggle with high data requirements, poor calibration (overconfidence), and an inability to quantify uncertainty. This paper introduces the Bayesian Topological Convolutional Neural Network (BTCNN), a hybrid architecture that integrates topological inductive biases with Bayesian inference. By leveraging the geometric structure of natural images through specialized topological layers and employing variational inference for classification, the BTCNN achieves faster convergence, superior performance in low-data regimes, and enhanced uncertainty quantification compared to traditional CNN architectures.

Introduction

Convolutional Neural Networks (CNNs) have revolutionized image processing but possess significant limitations: they are notoriously "data-hungry," often provide overconfident predictions even when incorrect, and lack a native mechanism to signal out-of-distribution (OOD) data. While Topological CNNs (TCNNs) use geometric priors to improve training efficiency and Bayesian Neural Networks (BNNs) offer robust uncertainty metrics, neither alone addresses all three issues effectively. The BTCNN aims to bridge this gap by combining the structural efficiency of topological manifolds with the probabilistic rigor of Bayesian inference.

Architecture and Methodology

The BTCNN architecture follows a "best of both worlds" design, splitting the network into a topological front end and a Bayesian back end.

Topological Feature Extraction

The initial layers utilize a Circle Filter (CF) and Circle-One Layer (COL). These components are grounded in the observation that patches of natural images tend to cluster around a "primary circle" (a Klein bottle embedding). By using fixed or constrained weights to encode these structural invariances, the model reduces the number of parameters that must be learned from scratch, effectively providing a geometric shortcut for feature detection.

The BTCNN architecture integrating topological layers (Circle Filter, Circle-One Layer) with Bayesian fully connected layers.

Bayesian Classification and Consistency

The fully connected layers at the back end employ Variational Inference (VI). Instead of static weights, the model learns probability distributions (mean and variance), allowing it to sample different weight configurations during inference to estimate predictive uncertainty. Additionally, the authors introduce a Consistency Condition (CC) term to the loss function. This regularizer penalizes the network if it produces divergent outputs for highly similar input images, encouraging smoothness and robustness against small perturbations.

Experimental Evaluation

The model was benchmarked using the USPS Dataset (9,298 grayscale images of handwritten digits). To evaluate robustness, the researchers tested the architecture under two challenging conditions:

  1. Data Starvation: Training on reduced subsets (25%, 50%, and 75%) of the data.
  2. Corrupted Inputs: Applying class-correlated Gaussian blur to test robustness against noise.

Results and Discussion

Faster Convergence and Efficiency

Experimental results indicate that the BTCNN reaches optimal performance significantly faster than standard models. Bayesian variants, including the BTCNN, typically leveled off in test loss after approximately 10 epochs, whereas non-Bayesian models required 15-20 epochs to achieve similar stability.

Test set loss over training epochs, illustrating the faster convergence of Bayesian and topological hybrid models.

Performance in Low-Data Regimes

In scenarios where only 25% of the training data was available, the BTCNN variants consistently outperformed standard CNNs and TCNNs in both raw accuracy and calibration. The inclusion of the Consistency Condition was particularly effective in lowering the Expected Calibration Error (ECE), ensuring that the model's confidence scores closely matched its actual probability of being correct.

Uncertainty Quantification

The BTCNN demonstrated a superior ability to identify OOD data. When presented with "convex combinations" (blends) of different digits, the model showed a sharp spike in epistemic uncertainty, whereas standard CNNs often made high-confidence, incorrect guesses. In noise-heavy experiments, the BTCNN maintained the lowest Maximum Calibration Error (MCE), proving its reliability under corruption.

Conclusion and Future Work

The BTCNN represents a significant step toward more robust and interpretable image classification. While the current model successfully integrates Bayesian stochasticity into the final dense layers, the authors noted that applying such stochasticity to the topological layers themselves currently degrades performance—a phenomenon requiring further investigation. Future research will explore the use of different topological manifolds beyond the unit circle and investigate methods for automatically learning the optimal manifold structure during training.

Summary updated

BibTeX
@misc{dayton2025bayesian,
  author = {Sarah Harkins Dayton and Hayden Everett and Ioannis D. Schizas and David L. Boothe and Vasileios Maroulas},
  title = {{Bayesian Topological Convolutional Neural Nets}},
  year = {2025},
  eprint = {2510.11704},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.11704},
}
Figure