Bayesian Quantum Orthogonal Neural Networks for Anomaly Detection

Natansh Mathur, Brian Coyle, Nishant Jain, Snehal Raj, Akshat Tandon, Jasper Simon Krauser, Rainer Stoessel · 2025

Read Bayesian Quantum Orthogonal Neural Networks for Anomaly Detection by Natansh Mathur, Brian Coyle, Nishant Jain, Snehal Raj, Akshat Tandon, Jasper Simon Krauser, Rainer Stoessel (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-18103.txt.

Introduction

Natansh Mathur and six co-authors combine in 2025 Bayesian learning with orthogonal neural networks, in quantum and classical forms, to spot defects in 3D objects. Anomaly detection in manufactured parts guards quality, and knowing a model's uncertainty matters as much as its verdict. Orthogonal weight matrices keep training smooth, while Bayesian layers price in doubt.

The team built orthogonal versions of 3D convolutional networks and trained them on an industrial test case. The models flagged anomalies well. They also ran versions on IBM's 127-qubit Brisbane quantum device to probe noise and shot limits. The hardware trials mark an early, honest step toward quantum help in factory inspection.

Catalog record

CREATORNatansh Mathur, Brian Coyle, Nishant Jain, Snehal Raj, Akshat Tandon, Jasper Simon Krauser, Rainer Stoessel
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-18103

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Bayesian Quantum Orthogonal Neural Networks for Anomaly Detection Natansh Mathur, Brian Coyle, Nishant Jain, Snehal Raj, Akshat Tandon, Jasper Simon Krauser, Rainer Stoessel Abstract Identification of defects or anomalies in 3D objects is a crucial task to ensure correct functionality. In this work, we combine Bayesian learning with recent developments in quantum and quantum-inspired machine learning, specifically orthogonal neural networks, to tackle this anomaly detection problem for an industrially relevant use case. Bayesian learning enables uncertainty quantification of predictions, while orthogonality in weight matrices enables smooth training. We develop orthogonal (quantum) versions of 3D convolutional neural networks and show that these models can successfully detect anomalies in 3D objects. To test the feasibility of incorporating quantum computers into a quantum-enhanced anomaly detection pipeline, we perform hardware experiments with our models on IBM’s 127-qubit Brisbane device, testing the effect of noise and limited measurement shots. 1 Introduction In recent years, the intersection of quantum computing and machine learning has fostered the development of novel computational paradigms that push the boundaries of traditional algorithms [ DTH+25 ] . Among these, certain families of quantum machine learning (QML) models are efficiently simulatable by classical computers. Such quantum- inspired models have garnered significant interest, as they allow quantum machine learning solutions to be developed which do not require quantum computers to be fully-fledged but which can benefit from them when they are. Foremost among these are tensor networks, which have demonstrated the ability to successfully solve machine learning tasks [ SS16 ] . An alternative family are quantum-inspired parametrisations of orthogonal neural networks [ LML+22 , CKM+22 ] . Orthogonal neural networks (OrthoNNs) are a class of machine learning models that enforce orthogonality constraints on weight matrices. They have emerged as a promising tool due to their inherent stability, reduced risk of overfitting, and improved generalisation capabilities. The quantum-inspired versions of these networks inherit these positive features, but also can be implemented in a fully quantum mode, potentially delivering speedups for orthogonal neural networks at larger scales. In this work, we demonstrate the effectiveness of the Bayesian learning paradigm to train these quantum and quantum-inspired OrthoNN parametrisations. Bayesian learning is a probabilistic framework that facilitates principled uncertainty quantification and robust decision-making. By combining the structural advantages of OrthoNNs with the probabilistic strengths of Bayesian inference, we introduce Bayesian (quantum) orthogonal neural networks (BONNs) as a framework for machine learning that ensures efficient training while providing uncertainty-aware predictions. As an example application, we apply these BONNs to anomaly detection, a critical task in various industrial domains such as manufacturing, energy systems, and cybersecurity. Detecting anomalies involves identifying patterns or behaviours that deviate significantly from normal operations, which is essential for ensuring safety, efficiency, and reliability. Traditional anomaly detection methods often struggle with noisy data, complex patterns, and the need for high interpretability—challenges that BONNs are uniquely positioned to address. We explore the efficacy of BONNs in anomaly detection through an industrially relevant use case, demonstrating their ability to identify anomalies with high precision and confidence. By focusing on anomalies in three-dimensional (3D) objects, we also incorporate orthogonal layers into 3D convolutional neural networks and propose the 3D orthogonal convolutional NN (OrthoConv3D). We compare the efficacy of this architecture to basic (orthogonal) feedforward networks. 2 Background Figure 1 : Overview of the proposed anomaly detection pipeline using orthogonal neural networks. The 3D object (data point) is passed through 3D convolutional layers of the autoencoder, which are parametrised with orthogonal weight matrices. In the autoencoder framework, an Encoder compresses information to a 3D latent space, and a Decoder attempts to reconstruct the image from it. If the reconstruction is deemed successful, the object contains no anomaly. In the orthogonal 3D convolution, 3D patches (red) of a tensor object are flattened and treated as feature vectors. The 3D CNN kernels (grey) are flattened and form a matrix modelled by ortholinear quantum circuits (or quantum-inspired simulation) and multiplied with the flattened patches. The Loader circuit encoding this feature vector depicted is the ‘diagonal’ loader [ JDM+21 ] . The orthogonal matrix multiplication corresponds to flattened 3D kernels. The orthogonal circuit shown is the ‘pyramid’ structure [ LML+22 ] . In the Bayesian framework, each orthogonal circuit parameter is sampled from a distribution, e.g. Gaussian with trainable mean and variance. 2.1 Additive manufacturing Additive Manufacturing (AM) (also known as 3D printing) techniques have the ability to create and produce complex geometries through optimised design approaches. The potential for mass and cost savings has made AM highly competitive compared to conventional manufacturing processes, driving its rapid adoption in fields such as aerospace and automotive [ HBT+23 , GRSK21 ] . 2.2 Traditional techniques for anomaly detection in additive manufacturing Despite extensive research and technological advancements in AM processes, identifying internal defects (anomalies) remains a persistent challenge [ HBT+23 ] . These defects can arise from various factors, including non-uniform powder distribution, variations in process parameters affecting the laser beam, scanning and building strategies, and deformations during the manufacturing process [ GKHR18 , YS17 ] . Common defects in AM parts include inclusions, pores, and lack of fusion (LoF). Inclusions result from foreign material contamination, while pores and LoF are typically caused by process-related issues such as gas bubbles trapped during melt pool solidification and insufficient energy from the melting beam, respectively. These defects significantly impact the mechanical properties and overall quality of the 3D-printed components [ GPSM17a , CME+22b , GPSM17b , CME+22a ] A common family of approaches to identify, analyse, and classify defects are non-destructive testing (NDT) methods. These methods enable manufacturers to determine whether a part meets predefined industrial acceptance criteria based on defect size, shape, type, and distribution patterns. Among the various NDT methods, computed tomography (CT) is one of the most widely used due to its ability to provide detailed information about both the outer geometrical features and the internal structure of the part. CT scans offer a three-dimensional view, enabling comprehensive defect analysis [ BCB+25b , MDH24 , BCB+25a ] . However, traditional CT inspection processes often involve manual analysis by human experts, which can be time-consuming, labour-intensive, and prone to inaccuracies [ KSBS18b , KSBS18a ] . 2.3 Machine learning for anomaly detection in additive manufacturing To address these challenges, computer vision-based methods have recently been applied to automate defect detection in CT scans [ OSR08 , CRS+97 ] , which can identify defects more efficiently and consistently. Using such automated approaches has shown promise post-printing, though their performance is often hindered by artefacts inherent to CT imaging, such as beam hardening, scatter, Poisson noise, and ringing effects and are often both time and cost intensive [ BF+12 ] . Besides the cost and time constraints, the adoption of additive manufacturing (AM) for critical aviation components is also limited by the capabilities of computed tomography (CT) machines. These machines can only inspect objects within a specific size range, restricting the geometry and dimensions of AM parts that can be effectively screened. As a result, this limitation has impeded the broader integration of AM components in aviation applications [ MDH24 , Hol25 ] . Therefore, to broaden the application of AM parts in critical structures, especially in aviation sectors, while simultaneously reducing production and inspection

Read the complete text