When a parameterized quantum circuit (a quantum program whose gate angles are trained like neural-network weights) learns a sequence of tasks, the standard way to stop it from forgetting is Elastic Weight Consolidation (EWC). EWC adds a penalty that pulls each angle back toward its old value, weighted by how important that angle was for earlier tasks. Importance is estimated as the average squared gradient, which is the diagonal of the Fisher information. On real quantum hardware every gradient is estimated from a finite number of measurements, or "shots," so every squared gradient carries a positive noise term that does not average out. Our claim is that this term adds a near-uniform floor to every importance weight, and that on noisy devices the floor dominates. EWC on hardware therefore behaves like isotropic weight decay toward the old parameters: it holds everything equally still. It becomes rigid where it should be flexible, and it fails to single out the few angles that actually matter.
This links the simulation-to-hardware gap to a known failure mode of continual learning. Classical theory already says that uniform penalties cannot reconcile forgetting and the ability to learn new tasks when tasks differ. We propose to show that hardware noise quietly moves quantum EWC into exactly that losing regime. We also propose a cheap fix: an unbiased Fisher estimator built from independent shot batches. It removes the floor at no extra shot cost, and we expect it to recover most of the performance of an oracle that knows the true noisy-device Fisher.
Why this gap is real and not already closed
The nearest quantum continual-learning work regularizes the wrong object for this question, or works without noise. Continual Quantum Architecture Search with Tensor-Train Encoding runs on IBM's Heron r2 processor, but its EWC anchors the classical Transformer agent that proposes circuit layouts, not the gate angles executed on the chip. Universal Adversarial Perturbations for Multiple Classification Tasks with Quantum Classifiers does apply EWC to circuit angles across MNIST, MedNIST and quantum phase-classification tasks. It does so in noiseless state-vector simulation and treats EWC as a black box, so it never measures rigidity. Quantum Architecture Search via Continual Reinforcement Learning handles drifting noise by reusing stored policies rather than by any parameter penalty, on a two-qubit Bell-state task in simulation. None of these asks how the importance estimate itself behaves when it is measured through shots on a noisy device.
Two classical results supply the lens. On the Computation of the Fisher Information in Continual Learning shows that biased shortcuts for the diagonal Fisher (the "empirical" and "batched" squared-gradient estimators) measurably hurt EWC. On Split CIFAR-10 the exact estimator reaches about 84.9% versus 83.28% for the empirical one, and the biased estimators shift the best penalty strength substantially. Shot noise is a far larger version of this kind of bias. Fixed Design Analysis of Regularization-Based Continual Learning proves, for two-task linear regression, that an isotropic $\ell_2$ penalty trades forgetting against intransigence (failure to learn the new task) and cannot match joint training on highly dissimilar tasks at any penalty strength. That result constrains the design: if the shot-noise floor makes EWC isotropic, no amount of tuning the penalty strength will rescue it.
The hardware side motivates why the floor grows with device noise. RoQNN: Noise-Aware Training for Robust Quantum Neural Networks finds that hardware noise acts on measured expectation values roughly as an affine map, shrinking them by a factor $\alpha<1$ and shifting them. Its fix, post-measurement normalization, divides that shrinkage back out. Normalization restores the scale of the signal but also multiplies the shot noise by $1/\alpha$. So the very step that makes noisy circuits trainable makes the Fisher floor worse.
Core Claim
Let $g_i$ be the true gradient of the per-example log-likelihood with respect to angle $i$ on the noisy device, estimated by the parameter-shift rule (evaluating the circuit at shifted angles) from $N$ shots per evaluation. Let $\sigma_i^2$ be the single-shot variance of that estimate after post-measurement normalization. The squared-gradient estimator then satisfies
$$\mathbb{E}[\hat g_i^{\,2}] = g_i^2 + \frac{\sigma_i^2}{N}, \qquad \sigma_i^2 \propto \frac{1}{\alpha^2}.$$
The first term is the real importance. The second term is roughly the same for every angle, because parameter-shift variance depends mostly on the measured observable rather than on which angle is shifted. It also grows as the device's contraction factor $\alpha$ falls. When it dominates, the EWC penalty $\tfrac{\lambda}{2}\sum_i \Omega_i(\theta_i-\theta_{\text{old},i})^2$ approaches $\tfrac{\lambda c}{2}\|\theta-\theta_{\text{old}}\|^2$ for some constant $c$.
Our fix draws two independent shot batches $a$ and $b$ for each parameter-shift evaluation, the same total budget as one batch of size $N$. It estimates importance as
$$\hat\Omega_i = \max\!\big(0,\ \overline{\hat g_i^{(a)}\hat g_i^{(b)}}\big),$$
where the bar averages over examples. Because the two batches are independent, the product has expectation exactly $g_i^2$, with no floor. Following the classical finding that sampling labels from the model's own predictions beats using ground-truth labels, we compute $g_i$ with labels sampled from the circuit's normalized predictive distribution. The method changes nothing else about EWC, and it is compatible with RoQNN-style noise-injected training.
The claim is falsifiable on three counts. On noisy devices, standard quantum EWC should show an anisotropy (ratio of largest to median importance) far below that of the true noisy-device Fisher. Its forgetting-versus-intransigence curve should lie on top of the curve for isotropic $\ell_2$. The two-batch estimator should move that curve toward the oracle's.
Research Program
The first stage is diagnostic. In density-matrix simulation with noise models built from IBM calibration data, we can compute the exact noisy-device Fisher with infinite shots. This shows directly how much of each estimated importance is floor rather than signal, as a function of shot count, of an artificial noise scale that sweeps $\alpha$, and of normalization being on or off. We will also extend the two-task linear analysis of Li et al. to a penalty of the form $\Omega + cI$. The goal is a closed-form statement of how the intransigence ceiling rises with $c$, which gives a theoretical prediction to check against the circuits.
The second stage is continual learning in device-calibrated simulation. We will use small (4- to 8-qubit) classifiers of the kind RoQNN deploys, trained in TorchQuantum or Qiskit Aer. They will see task sequences of increasing dissimilarity: pairwise MNIST class splits, MNIST then Fashion-MNIST, and the MNIST-to-MedNIST pairing from the adversarial-perturbation paper. We will also run the 8-task MIT-BIH ECG stream used by CL-QAS, but with EWC applied to the circuit angles themselves. A drift condition will swap the calibration snapshot between tasks, so that importance is measured under one noise profile and the penalty is enforced under another.
The third stage is a limited confirmation on physical IBM hardware. Because physical runs are expensive and unstable, as Benchmarking Quantum Reinforcement Learning documents, we restrict this stage to one two-task sequence. In it we estimate the Fisher on the device using standard and two-batch estimators at an equal shot budget, train task two in simulation under the same calibration, and evaluate both tasks on hardware.
Baselines
The naive reference is plain sequential fine-tuning with no penalty, which measures maximal forgetting. Its counterpart is a circuit trained jointly on all tasks, which gives the accuracy ceiling and the per-task reference used to define intransigence. Isotropic $\ell_2$ anchoring to the previous angles is the baseline our claim says standard EWC collapses into. It is tuned over the same penalty-strength grid as every EWC variant.
The strongest existing competitor is standard EWC on circuit angles with the squared-gradient Fisher, as used in Universal Adversarial Perturbations for Multiple Classification Tasks with Quantum Classifiers, here combined with RoQNN normalization and noise injection so that it is not handicapped on noisy devices. That paper's merged clean accuracy of 93.3% on MNIST and MedNIST is a noiseless reference point only. We found no reported result on Lacuna for EWC applied to circuit angles under hardware noise, so we will establish that number ourselves. For hardware context, RoQNN reports 80% four-class MNIST accuracy on real IBM devices. CL-QAS reports 86.3% accuracy on the ECG stream on Heron r2, but it is an architecture-search system rather than angle-level EWC, so we cite it as context, not as a like-for-like competitor.
Every method gets the same circuits, data order, total shot budget for Fisher estimation and training, and penalty-strength grid. The upper reference is EWC using the exact noisy-device Fisher. The ablations are two-batch debiasing alone versus label sampling alone, which isolates which ingredient carries the gain. We count it as beating standard EWC if, at equal shots, the two-batch estimator closes at least half the gap to the oracle-Fisher EWC on average accuracy. It must also reduce intransigence without raising forgetting, across the full tuning sweep and not just at one penalty strength.
Why not just…?
Why not just tune the penalty strength on hardware? Retuning can correct an overall scale error, but the problem we predict is one of shape. A floor-dominated importance vector is nearly uniform, and Li et al. prove that a uniform penalty has a forgetting-intransigence ceiling that no strength escapes on dissimilar tasks. We will test this directly by plotting the full forgetting-versus-intransigence frontier across the penalty grid for each estimator. If standard EWC's frontier matches the two-batch frontier, our idea is wrong and tuning suffices.
Why not just spend more shots on the Fisher? The floor does shrink as $1/N$, so this is a genuine competitor. Because the floor also grows as $1/\alpha^2$, however, the required $N$ rises steeply on noisier devices, and shots are the scarce resource on hardware. We will compare standard EWC at 2×, 4× and 8× the shot budget against the two-batch estimator at 1×. This shows how many shots buy the same frontier. If modest multiples close the gap, the contribution shrinks to a cost saving and we will report it as such.
Why not just rely on RoQNN's normalization and noise injection? Normalization undoes the shrinkage of the signal but, as argued above, amplifies shot noise by the same factor. Noise injection pushes the circuit toward flat regions where true gradients are small, which further lowers signal relative to the floor. Running both with and without these techniques under each estimator will show whether they reduce or worsen floor dominance.
Why not just retrain from scratch, or use one circuit per task? Retraining needs stored data, which streaming ECG and medical settings often forbid, and repeated hardware training is costly and unstable Benchmarking Quantum Reinforcement Learning. Separate circuits multiply qubit and calibration demands and give up transfer between tasks. Joint training stays in the study as the accuracy ceiling rather than as a competitor.
Proposed Experiments and Decisive Checks
The first decisive check is the diagnostic. If, at shot counts typical of hardware experiments, the floor term accounts for well under half of the estimated importance on calibrated noise models, the premise fails and we stop. The second check is whether the anisotropy of standard EWC's importance falls toward one as the simulated noise scale rises, while the oracle's anisotropy stays high. The third is the frontier comparison on dissimilar sequences, MNIST to MedNIST and the ECG stream, where theory predicts the largest separation between isotropic and anisotropic penalties. On near-identical pairwise MNIST splits we expect little difference, which serves as a built-in control.
Metrics are average final accuracy, forgetting (drop on earlier tasks), intransigence (gap to the single-task reference on the new task), and Spearman correlation between estimated and oracle importance. All are reported over at least five seeds. The hardware run succeeds if the importance anisotropy measured on the device, and the ordering of the two estimators, match the simulation predictions.
References
- Continual Quantum Architecture Search with Tensor-Train Encoding: Theory and Applications to Signal Processing
- Universal adversarial perturbations for multiple classification tasks with quantum classifiers
- Quantum Architecture Search via Continual Reinforcement Learning
- On the Computation of the Fisher Information in Continual Learning
- Fixed Design Analysis of Regularization-Based Continual Learning
- QuantumNAT: Quantum Noise-Aware Training with Noise Injection, Quantization and Normalization
- Benchmarking Quantum Reinforcement Learning