A hardware efficient quantum residual neural network without post-selection

Abstract: We propose a hardware efficient quantum residual neural network which implements residual connections through a deterministic linear combination of identity and variational unitaries, enabling fully differentiable training. In contrast to the previous implementation of residual connections, our architecture avoids post-selection while preserving residual learning. Furthermore, we establish trainability of our model, mitigating barren plateaus which are considered as a major limitation of variational quantum learning models. In order to show the working of our model, we report its application to image classification tasks by training it for MNIST, CIFAR, and SARFish datasets, achieving accuracies of 99% and 80% for binary and multi-class classifications, respectively. These accuracies are comparable to previously achieved from the standard variational models, however our model requires 10x fewer gates making it better suited for resource constraint near-term quantum processors. In addition to high accuracies, the proposed architecture also demonstrates adversarial robustness which is another desirable parameter for quantum machine learning models. Overall our architecture offers a new pathway for developing accurate, robust, trainable and hardware efficient quantum machine learning models.

• The paper introduces a differentiable quantum residual network that uses ancilla-controlled linear combinations of unitaries to remove post-selection while retaining residual learning and trainable layer strengths.
• The proposed five-block model achieves 99% accuracy on binary MNIST with 200 gates, while a hybrid QVC–QResNet reaches 80% on ten-class MNIST using 1,400 gates versus 8,000 for a deeper QVC.
• The paper shows that scaling the residual-strength bound as βmax = √d can prevent exponential gradient suppression under unitary-design assumptions, although hardware noise and real-device performance remain untested.

Overview

This paper by Khatun, Karim, and Usman (CSIRO/Data61 and the University of Melbourne) introduces a quantum residual neural network (QResNet) that implements residual connections through a deterministic linear combination of unitaries (LCU), eliminating the post-selection step required in prior residual quantum architectures. The central claims are threefold: the architecture is fully differentiable end-to-end, it provably avoids barren plateaus when the residual strength parameters are trained with appropriately scaled bounds, and it achieves classification accuracies comparable to standard variational quantum classifiers (QVCs) while using roughly an order of magnitude fewer gates. All experiments are classical state-vector simulations in PennyLane with a PyTorch backend; no hardware results are reported.

Architecture: LCU-based residual blocks without ancilla measurement

The model amplitude-encodes classical data onto $n$ data qubits and applies $L$ sequential residual blocks. In block $\ell$, an ancilla is prepared via $R_Y(\theta_\ell)$ into a coherent superposition; it controls whether the variational unitary $W_\ell(\vartheta_\ell)$ or the identity acts on the data register; then the ancilla is uncomputed with $R_Y(-\theta_\ell)$. The variational unitary consists of $R_z$–$R_y$–$R_z$ rotations on each data qubit followed by CNOT entanglers between neighboring qubits.

The key design choice is the ancilla preparation rule $\theta_l = 2\arctan(|\beta_l|)$, which yields amplitudes such that the unpostselected circuit realizes the deterministic map

$L$0

Because the raw expectation value $L$1 is automatically scaled by $L$2 per block, the network output

$L$3

is a smooth function of both the variational angles and the residual strengths. This contrasts sharply with the prior QResNet formulation of Heredge et al., where post-selection on the ancilla yields an effective map proportional to $L$4. Post-selection has two drawbacks the authors emphasize explicitly: acceptance probability decays exponentially with depth (the product of per-block success probabilities), and gradient information cannot propagate through a stochastic sample-and-discard step, breaking compatibility with gradient-based optimizers. By never measuring the ancilla, the proposed construction sidesteps both issues while retaining non-unitary behavior: Bloch-sphere analysis shows that whereas unitary QVC evolution preserves distances between input states, the residual map concentrates states along an axis—bifurcating toward antiparallel directions rather than collapsing to one axis as strict post-selection would.

A further departure from Heredge et al. is that each $L$5 is trainable rather than fixed. The limits are transparent: $L$6 bypasses the block entirely ($L$7), while $L$8 applies the equal mixture $L$9. Empirically, after training on binary MNIST the learned strengths converge to approximately $\ell$0, $\ell$1, $\ell$2, $\ell$3, $\ell$4—the optimizer learns to discard the fourth block outright. This adaptive layer-wise specialization is not available under fixed-$\ell$5 formulations.

For multi-class tasks, single-qubit expectation values across all data qubits form a logit vector scaled by the same normalization product, followed by softmax and cross-entropy loss. Because five residual blocks alone lack expressivity for ten-class MNIST, the authors prepend 30 standard QVC layers; the composite circuit remains fully differentiable.

Trainability: analytic barren plateau bounds

The paper derives gradient variance under Haar-random (unitary 2-design) assumptions using Weingarten calculus. For fixed $\ell$6, the variance scales as $\ell$7 and barren plateaus persist. However, treating $\ell$8 as a scalable hyperparameter changes this: setting $\ell$9 gives

$R_Y(\theta_\ell)$0

for gradients with respect to the variational angles, and

$R_Y(\theta_\ell)$1

for gradients with respect to the residual strengths. Both approach constants rather than decaying exponentially with system size, so any $R_Y(\theta_\ell)$2 guarantees absence of barren plateaus for arbitrary Haar-random unitaries. Notably, the appendix extends this to the hybrid QVC+QResNet architecture, showing that the residual term compensates for the exponential suppression present in the pure-QVC cost term—the combined variance retains the same favorable scaling. This is a stronger result than merely avoiding plateaus within the residual blocks themselves. The derivation assumes arbitrary input density matrices $R_Y(\theta_\ell)$3, though it relies on the standard unitary-design idealization; deviations under hardware noise or structured (non-Haar) ansätze are not analyzed.

Empirical results

All simulations use Adam (learning rate $R_Y(\theta_\ell)$4, weight decay $R_Y(\theta_\ell)$5). The headline comparison:

Model	Task	Test Acc. (%)	Total Gates
QResNet (5 blocks)	MNIST binary	99	200
QResNet (5 blocks)	CIFAR-2 binary	76	200
QResNet (5 blocks)	SARFish binary	72.14	200
QVC-200	MNIST 10-class	85	8000
QVC-30	MNIST 10-class	65	1200
QVC-30 + QResNet	MNIST 10-class	80	1400

Two numerical results stand out. First, binary MNIST exceeds 99% accuracy within the first few epochs using only 200 gates versus roughly 8000 for a deep QVC baseline—a gate reduction the abstract characterizes as 10×. Second, appending five residual blocks to a shallow 30-layer QVC lifts 10-class MNIST accuracy from 65% to 80% without increasing variational depth, adding only 200 gates. Since entangling gates dominate NISQ decoherence, this depth reduction bears directly on hardware feasibility. Performance degrades predictably on harder data: CIFAR-2 reaches 76% and SARFish 72.14%, demonstrating applicability to noisy remote-sensing data.

On adversarial robustness, evaluated with FGSM perturbations on 10-class MNIST, the model is vulnerable under white-box attacks (accuracy degrades with increasing $R_Y(\theta_\ell)$6, as expected for differentiable models) but remains largely stable under black-box attacks where adversarial examples are transferred from a classical neural network. The authors attribute this to decision boundaries structurally misaligned with classical models, consistent with earlier findings on quantum adversarial robustness.

Limitations and open questions

Several caveats should be weighed against these results. All experiments are noiseless state-vector simulations; the paper concedes that simulating deeper QResNet configurations was limited by classical memory constraints from the ancilla-controlled operations, and feasibility on physical hardware remains untested. The barren plateau guarantees hold under Haar-random/unitary-design assumptions and noiseless conditions, neither of which strictly obtains on NISQ devices. The multi-class architecture requires a 30-layer QVC backbone for sufficient expressivity, so the "5-blocks-only" efficiency claim applies cleanly only to binary tasks. The adversarial robustness claim rests on transfer attacks from a single classical surrogate; broader attack surfaces (e.g., quantum-native white-box optimization beyond FGSM) are not explored. Open questions include how the trainable-$R_Y(\theta_\ell)$7 mechanism behaves under realistic depolarizing noise, whether the $R_Y(\theta_\ell)$8 bound is achievable in practice given finite-precision parameter encoding, and how the architecture scales beyond 10 data qubits.

Conclusion

This paper presents a post-selection-free quantum residual architecture built from ancilla-controlled LCU blocks with trainable residual strengths, providing end-to-end differentiability, an analytically grounded barren plateau mitigation strategy, and competitive image-classification accuracy at substantially reduced gate counts. Its principal contribution is architectural: demonstrating that residual learning can be retained deterministically, with provable gradient scaling, rather than probabilistically. Validation on actual quantum hardware and under realistic noise remains the outstanding test of the approach.