⚛️ quantum physics

Invariance can prevent overfitting and shortcut learning in Quantum Neural Networks

This paper demonstrates that incorporating $D_4$ symmetry into quantum neural networks significantly improves generalization, robustness to shortcut learning, and noise resilience compared to non-invariant models with identical architectures and parameter counts across diverse image classification tasks.

Original author: Giuseppe D'Ambruoso, Giovanni Gramegna, Domenico Pomarico, Giuseppe Magnifico, Cosmo Lupo

Published 2026-11-11

📖 4 min read🧠 Deep dive

CC BY 4.0

Original author: Giuseppe D'Ambruoso, Giovanni Gramegna, Domenico Pomarico, Giuseppe Magnifico, Cosmo Lupo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

In the world of artificial intelligence, computers are learning to see the world by scanning images and finding patterns. When a machine looks at a photograph of a cat, it does not care if the cat is sitting upright, lying down, or viewed from a different angle; the label "cat" remains the same. This ability to recognize that the core identity of an object stays constant even when its position changes is called invariance. For decades, engineers have built classical computer programs that hard-code this rule, teaching them to ignore rotation and reflection so they can learn faster and more accurately. Now, scientists are trying to build a new kind of computer that uses the strange laws of quantum physics to process information. These quantum machines are incredibly powerful in theory, but they are currently very small and fragile, prone to making mistakes when they try to learn from limited data. A major question has been whether these new quantum machines can also learn to ignore the position of an object to become better learners, or if their unique physics makes them too sensitive to such details.

A team of researchers set out to answer this by building four different quantum models to solve image classification tasks. They wanted to see if forcing a quantum computer to respect the symmetry of an image—meaning it treats a rotated image as the same object—would help it learn better. To do this fairly, they constructed four distinct models that were identical in every way except for how they handled symmetry. They used the same number of quantum bits, the same number of adjustable settings, and the same underlying mathematical structure. Two of these models were designed to be "invariant" by nature, meaning their internal rules forced them to ignore the orientation of the image. The other two were allowed to notice the orientation, treating a rotated image as a completely different input. The researchers tested these models on four different challenges: recognizing handwritten numbers, identifying simple geometric shapes, distinguishing between two types of galaxies, and spotting different shapes of red blood cells. In each case, the images appeared in random rotations, just as they might in the real world.

The results were clear and consistent across all four tasks. The models that were built to ignore the orientation of the images learned much faster and made fewer mistakes on new, unseen pictures. Because they were forced to focus only on the shape of the object and not its position, they did not waste their limited learning capacity memorizing specific details of the training photos. In contrast, the models that were allowed to notice the orientation quickly memorized the exact poses of the few images they were shown. When these non-invariant models faced a new image with a different rotation, they often failed completely. The researchers found that simply making the circuit of the quantum computer symmetric was not enough; the final step where the computer reads out its answer also had to be symmetric. If the computer was built with symmetric rules but asked to give a non-symmetric answer, it still failed to generalize.

The study also explored a common problem in machine learning called "shortcut learning," where a model finds an easy trick to guess the answer instead of learning the real rule. The researchers added a small, hidden watermark to the corner of every training image to see if the models would rely on it. The models that could notice orientation learned this trick instantly; they would look at the corner, see the watermark, and guess the answer without ever looking at the main image. When the researchers moved the watermark to a different corner, these models failed, dropping to the level of random guessing. The invariant models, however, could not see the watermark at all because it moved with the image rotation. Instead, they were forced to look at the actual content of the picture, and they continued to identify the objects correctly even when the trick was removed.

Finally, the team tested how these findings held up against the noise and errors typical of current quantum hardware. Real quantum computers are imperfect, and their signals often get distorted by heat and interference. The researchers simulated these conditions and found that the advantage of the invariant models survived. Even when the quantum signals were noisy, the models that ignored orientation remained more accurate and less prone to overfitting than their counterparts. The study suggests that for quantum computers to be useful in the real world, especially when data is scarce, building symmetry directly into their design is a powerful tool. It acts as a guardrail, preventing the machine from getting distracted by irrelevant details and helping it focus on the true patterns that define the world around it.

Technical Summary: Invariance in Quantum Neural Networks

Problem Statement
Variational Quantum Algorithms (VQAs), specifically Quantum Neural Networks (QNNs), are a leading approach for machine learning on near-term quantum hardware. However, their utility is constrained by the limited number of qubits, noise, and the "barren plateau" problem, where gradients vanish exponentially with system size. Consequently, QNNs are typically trained on small datasets, raising a critical question: do they generalize effectively from few examples? While symmetry-aware architectures (equivariant networks) are theoretically expected to improve generalization and data efficiency by restricting the model to invariant functions, previous studies have struggled to isolate the specific benefit of symmetry. Often, equivariant models are compared against non-equivariant baselines that differ in gate sets, circuit structure, or parameter counts, making it unclear whether performance gains stem from symmetry or other architectural differences. Furthermore, it remains unclear whether the benefit arises from the equivariant circuit itself or the invariant readout, and whether these models are robust to "shortcut learning" (relying on spurious cues) and realistic hardware noise.

Methodology
The authors conduct a controlled investigation into the role of invariance in image classification tasks where labels are invariant under the dihedral group $D_4$ (rotations by multiples of $90^\circ$ and reflections). To isolate the effect of symmetry, they design a study comparing four distinct 8-qubit models that share:

Identical Encoding: Amplitude encoding of $16 \times 16$ pixel images onto 8 qubits, where $D_4$ acts via bit-flips and register swaps.
Identical Circuit Structure: The same number of layers, trainable parameters, and parameter-sharing patterns.
Equivalent Dynamical Lie Algebras (DLA): The circuits are constructed such that their DLAs are conjugate and of the same dimension ($569$), ensuring they possess the same expressive power regarding reachable unitaries.

The four models are defined by the combination of circuit equivariance and readout invariance:

Equiv: An equivariant circuit measured with an equivariant observable (fully invariant model).
Equiv-NoSym: The same equivariant circuit measured with a non-equivariant observable (non-invariant model).
NonEquiv: A non-equivariant circuit (constructed by conjugating the equivariant circuit with a fixed unitary on one register) measured with an equivariant observable (non-invariant model).
NonEquiv-Twirled: The non-equivariant circuit averaged over the group $D_4$ (classical post-processing) to enforce invariance.

The models are benchmarked on four binary classification tasks: MNIST digits (4 vs. 5), a synthetic task (cross vs. square), Galaxy10 (face-on vs. edge-on galaxies), and red blood cells (sickle-shaped vs. other deformed). The study evaluates generalization error, robustness to shortcut learning (using class-dependent watermarks and gradients), and resilience to a realistic noise model simulating superconducting quantum processors.

Key Contributions and Results
The study yields six primary findings:

Invariance Prevents Overfitting: On all four tasks, invariant models (Equiv and NonEquiv-Twirled) trained on as few as 20–40 images exhibit significantly smaller generalization gaps and higher test accuracy compared to non-invariant models. Non-invariant models tend to overfit the specific poses present in the training set.
Circuit Equivariance is Insufficient: The equivariant circuit alone does not guarantee generalization. The Equiv-NoSym model (equivariant circuit, non-invariant readout) overfits to the same extent as the non-equivariant NonEquiv model. The benefit arises only when the entire model (circuit + readout) is invariant.
Cost of Learning Symmetry: Non-invariant models overfit because they must learn the symmetry from the data. When tested on upright digits (where orientation is fixed and informative), the non-invariant model overfits significantly less, confirming that the overfitting in random-pose settings is the cost of inferring the symmetry constraint.
Built-in vs. Averaged Invariance: Both built-in invariance (Equiv) and classical group averaging (NonEquiv-Twirled) prevent overfitting. However, their accuracy depends on the task. Equiv performs better on the synthetic task, while NonEquiv-Twirled performs better on red blood cells. This is attributed to their different function classes: Equiv has a smaller, fully invariant feature space, while NonEquiv-Twirled projects a larger, non-invariant space onto invariant features.
Immunity to Orientation Shortcuts: When trained with class-dependent corner watermarks or brightness gradients (shortcuts), non-invariant models learn these spurious cues and fail when the cues are moved or removed. Invariant models, unable to distinguish between rotated versions of the image, ignore the position-dependent shortcuts and learn the actual content of the images.
Resilience to Noise: The performance advantage of invariant models persists under realistic noise models (thermal relaxation, depolarizing noise, readout errors) typical of current superconducting processors. While noise slightly breaks the symmetry of the Equiv model, the degradation is minimal (<2%), and the generalization gap remains favorable compared to non-invariant models.

Significance
The paper demonstrates that incorporating task symmetries into QNNs is a viable strategy to mitigate overfitting and shortcut learning in the data-limited regime of near-term quantum devices. Crucially, the study establishes that the benefit is not inherent to the equivariant circuit structure alone but requires the model output to be invariant. By controlling for circuit depth, parameter count, and dynamical Lie algebra, the authors show that invariance acts as a powerful inductive bias, allowing models to generalize from few examples without relying on spurious correlations. These results support the use of symmetry constraints in quantum model design, particularly when labeled data is scarce, while highlighting that the specific choice of invariant function class (built-in vs. twirled) must be tailored to the specific task.

Technical Summary: Invariance in Quantum Neural Networks

More like this