arXiv:2609.31066
Modeling quantum neural network gradient with reinforcement learning
Nhan Trong Luu, Duong Trung Luu, Nam Ngoc Pham, Thang Cong Truong
Abstract
Training quantum neural networks (QNNs) on near-term hardware remains hampered by two compounding difficulties: the exponential vanishing of gradient variance known as the barren plateau, and the $\mathcal{O}(L \cdot 2^n)$ time and memory cost of differentiating through an $n$-qubit, $L$-layer circuit. We propose RLQ-Grad, a reinforcement-learning-based optimizer in which a classical policy $π_φ$ (a spectrally-normalized PPO agent) learns to propose parameter updates directly, conditioned on the QNN's current parameters, loss, accuracy, and previous update. Because the surrogate gradient is emitted by a classical network rather than obtained by differentiating through the unitary $U(θ)$, its variance is not constrained by the barren plateau concentration bound, and its cost scales with the number of trainable parameters rather than the Hilbert-space dimension. We prove these properties formally and verify them on a hardware-efficient ansatz across four supervised benchmarks with up to $n=20$ qubits. RLQ-Grad preserves a near-flat gradient-variance curve where backpropagation, parameter-shift, and adjoint differentiation decay by 1 to 2 orders of magnitude. Accounting for the full training pipeline (PPO rollouts, actor-critic updates, and optimizer states), RLQ-Grad needs under 2 MB of memory and runs $2490\times$, $7876\times$, and $673\times$ faster per iteration than these three methods at $n=20$. It improves top-1 accuracy by up to $+10\%$ over gradient-based baselines on circuits of up to 12 qubits, and matches dedicated barren plateau mitigation methods on CIFAR-10 at 14 to 20 qubits, where evolutionary and gradient-free optimizers collapse to chance.
AI Summary
The paper asks whether a learned classical optimizer can reduce differentiation costs and improve supervised quantum neural network (QNN) training when circuit gradients become small. RLQ-Grad uses a PPO actor–critic with spectral normalization to propose circuit-parameter updates from current parameters, minibatch loss and accuracy, and the previous update. Classical input/output layers are trained by backpropagation; agents are trained separately for each dataset and circuit configuration. Experiments use noiseless statevector simulations of hardware-efficient ansätze on Breast Cancer, MNIST, Fashion-MNIST, and CIFAR-10. Across 2–12 qubits, reported validation accuracies exceed differentiation-based baselines, reaching 95.27% on MNIST and 87.62% on Fashion-MNIST at 12 qubits. At 20 qubits and 600 circuit parameters, CIFAR-10 accuracy is $53.02\pm2.83\%$, versus $40.13\pm2.91\%$ for backpropagation and $53.16\pm3.65\%$ for layerwise training. Resource benchmarks report 1.95 MB for the agent pipeline and CPU iteration speedups of $2490\times$ over backpropagation and $7876\times$ over parameter-shift; the complexity comparison excludes common circuit-forward costs. The theoretical argument establishes that policy-generated updates need not obey quantum-gradient variance concentration bounds; it does not establish that nonvanishing updates are useful descent directions. The paper explicitly acknowledges that concentrated losses also suppress reward information, so RLQ-Grad cannot remove exponential shot requirements in fully developed barren plateaus. Other stated limitations include poor local minima, vanishing gradients in classical preprocessing, small-circuit agent overhead, and lack of hardware, finite-shot, or noisy-device tests. Adjoint differentiation remains more memory-efficient in the reported benchmarks.
Central claim may not hold: The barren-plateau theorems do not establish useful descent directions or efficient policy learning, and the paper itself acknowledges that concentrated quantum rewards retain the fundamental training obstruction. The purported end-to-end computational advantage omits necessary circuit forward evaluations from its complexity accounting, so Hilbert-space-independent policy computation does not establish Hilbert-space-independent training cost. The stated CIFAR-10 PPO rollout requires approximately 281 MB for float32 observations and actions alone at 20 qubits, contradicting the rollout-inclusive full-pipeline memory claim below 2 MB. Adjoint differentiation retains statevector-sized working buffers, so a constant buffer count does not justify O(1) additional memory in qubit count. The different barren-plateau variance slopes reported for mathematically equivalent exact differentiation methods are unexplained and do not validate the claimed bypass.
Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.
- Comments
- NeurIPS 2026 Main Track (Poster). OpenReview: https://openreview.net/forum?id=lfrvu8xfsh