Symmetry in quantum neural networks boosts accuracy on image tasks

Researchers at Politecnico di Bari and Università degli Studi di Bari Aldo Moro have isolated the benefit of symmetry in quantum neural networks by comparing four eight-qubit models with identical encoding, circuit structure and number of trainable parameters. The work addresses a longstanding challenge in the field, as previous attempts to assess symmetry’s impact often altered other network characteristics alongside it.

The team demonstrated that invariant models trained on as few as tens of images outperformed non-invariant ones on four binary classification tasks, including datasets of galaxies and red blood cells, and proved more resilient to noise. This invariance also prevents the models from learning irrelevant orientation cues, a common problem in machine learning. Dihedral Group D_4 Symmetry in Quantum Neural Networks Their work centers on the dihedral group D4, which describes transformations like rotations and reflections of square images, and how incorporating this symmetry impacts model accuracy. The researchers benchmarked invariant implementations against non-invariant ones using generalization error, robustness to shortcut learning and resilience to a realistic noise model. A key expectation driving this research is that restricting a model to invariant functions should improve generalization, particularly when training with limited data. Limiting the generators of the circuit, the fundamental building blocks of quantum computation, should mitigate the “barren plateau” phenomenon, a common obstacle in training quantum neural networks. The study’s focus on the dihedral group D4 stems from its relevance to image processing, as reflections and 90-degree rotations are frequent transformations in visual data. An invariant classifier, as defined in the study, assigns the same output to inputs related by these D4 transformations. The action of this group on quantum states is mathematically described by a unitary representation, which establishes a compatibility condition between how classical data is encoded into quantum states and how the symmetry group transforms those states. A linear operator within the quantum circuit is considered “(G,R_H)-equivariant” if it commutes with the unitary representation of the symmetry group, ensuring the circuit respects the inherent symmetries of the data. If both the circuit and the observable used to read out the result are equivariant, the entire model becomes invariant under the group’s transformations.

The team demonstrated that an equivariant quantum channel combined with an equivariant observable yields a model that consistently produces the same output for transformed inputs, mathematically expressed as f(RH(g)ρ = f(ρ) for all group elements g and quantum states ρ. Controlled Comparison of Equivariant and Non-Equivariant QNN Circuits A controlled experimental setup revealed that maintaining circuit structure and trainable parameters while varying symmetry directly impacts generalization performance in quantum neural networks. The study addressed a gap in previous research where comparisons between equivariant and non-equivariant quantum neural networks often involved differing gates, entangling patterns, or parameter counts. By fixing these variables, the team aimed to determine if symmetry itself, the property of remaining unchanged under transformations like rotations and reflections, could demonstrably improve performance. The models were constructed with amplitude encoding, where the dihedral group D4 acts on qubits via bit flips and register swaps, allowing for a clear distinction between equivariant and non-equivariant circuits. This setup enabled the creation of four distinct models: an equivariant circuit with an equivariant observable (Equiv), an equivariant circuit with a non-equivariant observable (Equiv-NoSym), a non-equivariant circuit (NonEquiv), and a non-equivariant circuit averaged over the D4 group (NonEquiv-Twirled). This advantage persisted even when the models were subjected to a realistic noise model, suggesting that invariance contributes to robustness. The researchers demonstrated that invariance effectively prevents models from falling prey to orientation shortcuts, a phenomenon where models learn to rely on irrelevant cues like the presence of a class-dependent corner watermark. When trained with such a watermark, non-invariant models often achieved chance-level performance when the watermark was moved, while invariant models remained unaffected. “A systematic comparison on several data sets, in which the circuit structure is fixed and the symmetry is varied in a controlled way, is still missing, and the effect of the symmetry itself remains unclear,” the paper states, highlighting the importance of this focused approach.

The team’s findings build upon established principles of quantum mechanics, specifically the concept of dynamical Lie algebras and their influence on circuit properties. They demonstrated that the effective observables of all circuits, built from the gates, span a space closed under commutators with the group’s generators. The researchers emphasize that both circuit equivariance and an equivariant observable are necessary for achieving invariance, as an equivariant circuit paired with a non-equivariant observable does not guarantee an invariant model. Impact of Invariance on Generalization and Shortcut Learning Researchers demonstrated that non-invariant models, when trained with a class-dependent corner watermark, frequently achieved chance-level performance when the watermark’s position shifted, a failure not observed in their invariant counterparts. This suggests a fundamental difference in how the models process visual information, with invariant networks focusing on core features rather than superficial details. This immunity to orientation shortcuts arises from a design choice that also improves generalization performance with limited data.

The team compared four eight-qubit models, each identical in encoding, circuit structure, and trainable parameters, but differing in their implementation of invariance. The benefits of invariance extend beyond simply achieving higher accuracy; it also addresses the issue of overfitting, a common challenge in machine learning. The study found that non-equivariant circuits, when classifying upright digits, overfit to a lesser extent than their equivariant counterparts, because they do not need to learn the symmetry from the images themselves. While both built-in invariance and classical post-processing methods like group twirling can prevent overfitting, the research highlights that the choice between them depends on the specific task at hand, with either approach potentially yielding greater accuracy. The work demonstrates that restricting the model to invariant functions improves generalization from limited examples, a valuable advantage in data-scarce environments. Geometric quantum machine learning principles underpin this improved performance, with the researchers noting that if a circuit is equivariant with respect to a symmetry group’s representation and the measured observable is also equivariant, the model output becomes invariant by construction. The study’s design focused on a controlled setting, classifying images invariant under the dihedral group D4, to isolate the specific benefit of symmetry.

Dynamical Lie Algebra and Circuit Parameter Consistency The models were benchmarked on their generalization error, resilience to shortcut learning and performance under realistic noise conditions. This rigorous approach aimed to isolate the benefit of invariance, a challenge because previous studies often altered multiple aspects of a network when introducing symmetry. A key finding is that the dynamical Lie algebra, generated by the circuit’s gates, dictates the range of unitaries the circuit can achieve and the number of parameters needed before it becomes overparametrized. These factors, in turn, influence the size of the gradients used during training. “Every circuit built from these gates lies in the Lie group eg,” the researchers write, describing the mathematical relationship between the gates and the achievable transformations. The researchers also explored how the choice of observable, the mathematical function used to extract information from the quantum circuit, interacts with symmetry. They combined equivariant (symmetry-respecting) and non-equivariant observables with both equivariant and non-equivariant circuits. This allowed them to assess the impact of an equivariant observable, which, when paired with an equivariant circuit, guarantees an invariant model output. Invariant models are immune to a specific type of shortcut learning. Consequently, the accuracy of non-invariant models plummeted when the watermark was moved or removed, while the invariant models remained unaffected, demonstrating their ability to focus on meaningful, symmetry-respecting features. The combination ensures that the model’s output remains consistent under transformations dictated by the symmetry group.

Invariance Resilience Under Realistic Noise Conditions Unlike non-invariant models, which quickly memorize specific images and their orientations when trained with only a few tens of examples, invariant models generalize more effectively to unseen data, suggesting a fundamental advantage in learning robust representations. This improved generalization isn’t simply a result of circuit design. The research demonstrates that circuit equivariance alone is insufficient, and the overall model, circuit combined with observable, must be invariant to prevent overfitting. An important distinction emerged when the models were challenged with a deliberately misleading “shortcut”, a class-dependent corner watermark added to the images. Non-invariant models readily learned to identify this artificial pattern, achieving high accuracy only when the watermark was present, and performance plummeted to chance levels when it was moved or removed.

The team’s investigation extended to a detailed analysis of how invariance is achieved within the quantum circuits, comparing four models each with eight qubits. These models differed in their implementation of symmetry, specifically examining combinations of equivariant and non-equivariant circuits paired with equivariant and non-equivariant observables. The results revealed that the function class, the set of all possible outputs the model can produce, is key to understanding performance. Further analysis confirmed that an equivariant circuit, when measured with a non-equivariant observable, overfits to the same degree as a completely non-equivariant circuit, highlighting the necessity of a fully invariant system. The researchers demonstrated this using mathematical principles, noting that a condition for invariance is more easily stated in the Heisenberg picture, where the adjoint of a transformation applied to a circuit reveals its symmetry properties. The study’s structure systematically explores these relationships, beginning with a review of equivariance, twirling and dynamical Lie algebras, and then detailing the encoding, circuits and models used in the experiments. The researchers also define the tasks, shortcuts, training procedures, and noise model employed to benchmark the models. The findings, they state, demonstrate that invariance is a theoretical benefit and a practical advantage in building robust and accurate quantum neural networks, even in the presence of realistic imperfections. 👉 More information🗞 Invariance can prevent overfitting and shortcut learning in Quantum Neural Networks✍️ Giovanni Gramegna, Domenico Pomarico and Giuseppe Magnifico (Dipartimento Interuniversitario di Fisica, Università degli Studi di Bari Aldo Moro, Bari, Italy); Cosmo Lupo (Dipartimento Interuniversitario di Fisica, Politecnico di Bari, Bari, Italy); Giuseppe D'Ambruoso🧠 ArXiv: https://arxiv.org/html/2611.14368v1 More like thisQuantum Machine LearningResearchers map qubit states to probability distributions in phase spaceHigh Performance ComputingGerman funds QUDORA’s quantum machine learning project for energyHigh Performance ComputingQuantum Inspire links to DelftBlue to test power grid calculationsQuantum Machine LearningQuantum Data Proposes DYRK1A Compounds Differing from Tabular ModelsStay currentSee today’s quantum computing news on Quantum Zeitgeist for the latest breakthroughs in qubits, hardware, algorithms, and industry deals. Tags: