Beyond Barren Plateaus: A Scalable Quantum Convolutional Architecture for High-Fidelity Image Classification

Radhakrishnan Delhibabu

cs.LGMar 11, 2026

Beyond Barren Plateaus: A Scalable Quantum Convolutional Architecture for High-Fidelity Image Classification

Authors: Radhakrishnan Delhibabu

Organizations: School of Computing Science and Engineering (SCOPE), Vellore Institute of Technology, Vellore, 632014, India.

Abstract

While Quantum Convolutional Neural Networks (QCNNs) offer a theoretical paradigm for quantum machine learning, their practical implementation is severely bottlenecked by barren plateaus -- the exponential vanishing of gradients -- and poor empirical accuracy compared to classical counterparts. In this work, we propose a novel QCNN architecture utilizing localized cost functions and a hardware-efficient tensor-network initialization strategy to provably mitigate barren plateaus. We evaluate our scalable QCNN on the MNIST dataset, demonstrating a significant performance leap. By resolving the gradient vanishing issue, our optimized QCNN achieves a classification accuracy of 98.7%, a substantial improvement over the baseline QCNN accuracy of 52.32% found in unmitigated models. Furthermore, we provide empirical evidence of a parameter-efficiency advantage, requiring $\mathcal{O}(\log N)$ fewer trainable parameters than equivalent classical CNNs to achieve $>95\%$ convergence. This work bridges the gap between theoretical quantum utility and practical application, providing a scalable framework for quantum computer vision tasks without succumbing to loss landscape concentration.

Figures & tables

Panel of Figure 1 — Figure 1: Conceptual visualization of the loss landscape. Standard QCNNs with global observables suffer from barren plateaus (right), where the gradient is exponentially close to zero everywhere except in a vanishingly small region near the minimum. Our proposed architecture yields a trainable landscape (left).

Panel of Figure 2 — Figure 2: Flowchart of the Proposed Scalable QCNN Architecture. The integration of Tensor Network Initialization and Localized Cost functions prevents the optimization process from stalling in a barren plateau.

Model Architecture	Parameter Complexity	Barren Plateau?	Max Qubits Simulated	Empirical Accuracy
Classical CNN (ResNet-lite)	$\mathcal{O}(N^{2})$	No (Vanishing Gradients fixed via ReLU)	N/A	$99.9\%$
Baseline QCNN (Global Cost)	$\mathcal{O}(\log N)$	Yes (Exponential decay)	12	$52.32\%$
Proposed Scalable QCNN	$\mathcal{O}(\log N)$	No (Provably avoided)	16	$\mathbf{98.7\%}$

Table 1: Theoretical and Empirical Comparison of Classical CNN, Baseline QCNN, and Proposed Scalable QCNN on the MNIST Dataset.

Gate Module	Operator Decomposition	Parameters
$RX$ Gate	$e^{-i(\theta/2)\sigma_{x}}$	1
$RY$ Gate	$e^{-i(\theta/2)\sigma_{y}}$	1
$RZ$ Gate	$e^{-i(\theta/2)\sigma_{z}}$	1
CZ Gate	$\|0\rangle\langle 0\|\otimes I+\|1\rangle\langle 1\|\otimes\sigma_{z}$	0
Conv Block	$R_{y}^{\otimes 2}\cdot\text{CZ}\cdot R_{x}^{\otimes 2}$	4
Pool Block	$\text{CNOT}\cdot(I\otimes R_{y}(\theta))\cdot\text{CNOT}$	1

Table 2: Taxonomy of Parameterized Quantum Gates utilized in the Convolutional and Pooling Blocks. The expressibility is defined by the number of independent rotational parameters per layer.

Panel of Figure 3 — Figure 3: Quantum Circuit Diagram of the proposed QCNN. The architecture demonstrates amplitude encoding followed by translationally invariant convolutional blocks $U_{block}$ and partial-trace pooling blocks $V_{pool}$ . Discarded qubits (dotted lines) denote tracing operations. Crucially, the final measurement computes localized observables $\langle Z_{i}\rangle$ rather than a global projection.

Panel of Figure 4 — Figure 4: System architecture integrating Cirq and TensorFlow Quantum. The computational graph encapsulates the quantum state simulation, allowing classical optimizers to seamlessly interact with the parameterized quantum layers via the parameter-shift rule.

Hyperparameter	Value / Specification
Qubits ( $n$ )	10
Amplitude Padding Length	$1024$
Initial Learning Rate ( $\eta_{0}$ )	$0.015$
Learning Rate Scheduler	Exponential Decay ( $\gamma=0.9$ )
Optimizer	Adam ( $\beta_{1}=0.9,\beta_{2}=0.999$ )
Batch Size ( $B$ )	32

Table 3: Hyperparameter Configuration for the TFQ-Cirq Simulation Pipeline.

Panel of Figure 5 — Figure 5: Empirical scaling of the gradient variance. The global cost function exhibits the hallmark exponential decay characteristic of barren plateaus ( $\mathcal{O}(2^{-n})$ ). Conversely, the localized cost function integrated into our QCNN architecture transitions to a polynomial decay curve ( $\Omega(1/\text{poly}(n))$ ), preserving gradient signals even at higher qubit counts.

Panel of Figure 6 — Figure 6: Convergence dynamics of the proposed QCNN on the MNIST dataset. (a) The classification accuracy climbs aggressively, plateauing at an empirical maximum of $98.7\%$ on the training set and $98.1\%$ on the validation set. The baseline QCNN is trapped at $\penalty\ 52\%$ . (b) The corresponding MSE loss function demonstrates smooth, non-oscillatory convergence, indicative of an appropriately tuned learning rate and a well-behaved gradient landscape.

Metric	Classical CNN	Baseline QCNN	Proposed QCNN
Cost Function	Cross-Entropy	Global ( $C_{G}$ )	Local ( $C_{L}$ )
Initialization	He Normal	Random Unif.	TTN Pre-train
Parameters	$\sim 1.2\times 10^{5}$	$45$	$\mathbf{45}$
Training Epochs	20	150	150
Test Accuracy	$99.9\%$	$52.32\%$	$\mathbf{98.7\%}$
Convergence State	Global Min	Barren Plateau	Global Min

Table 4: Final Comparative Analysis of Model Architectures.

Panel of Figure 7 — Figure 7: Accuracy degradation of the proposed QCNN architecture when subjected to symmetric depolarizing noise. The use of local observables and logarithmic circuit depth provides substantial inherent resilience, maintaining $>90\%$ accuracy even at a $1\%$ physical error rate.

Panel of Figure 8 — Figure 8: The end-to-end physical deployment pipeline. The logical circuit, seeded by the classical TNI protocol, undergoes hardware-aware transpilation to conform to the connectivity constraints of the physical processor before executing on the QPU.

Hardware Constraint	Impact on Standard QML	Solution in Proposed Architecture
Restricted Connectivity	High SWAP overhead required for global entanglement.	Tree-tensor pooling natively isolates spatial features; localized cost function reduces necessary routing.
Short Coherence Times ( $T_{1},T_{2}$ )	Deep circuits decohere, resulting in completely mixed states.	$\mathcal{O}(\log n)$ depth prevents $T_{1}/T_{2}$ decay from overpowering the feature extraction layers.
Readout Infidelity	Global multi-qubit measurement errors compound exponentially.	Localized single-qubit Pauli-Z expectations limit measurement error to individual operational margins.
Barren Plateaus	Optimizer stalls due to noise-induced gradient flattening.	Tensor Network Initialization places weights directly in the convergence funnel prior to QPU execution.

Table 5: Mapping of Physical NISQ Constraints to Proposed QCNN Architectural Solutions.

Explore similar work

A Novel Parallel QCNN Architecture with Efficient Classical Simulability
Jul 9, 2026Lawrence Nguyen, Hiu Yung WongQuantum Convolutional Neural Networks Quantum Circuit Simulation
Department of Electrical Engineering San Jos´e State University San Jos´e, USA
Layer-wise QUBO-Based Training of CNN Classifiers for Quantum Annealing
Mar 3, 2026Mostafa Atallah, Rebekah HerrmanImage Classification Hybrid Quantum-Classical ML
Department of Industrial and Systems Engineering, University of Tennessee - Knoxville, Knoxville, 37996, TN, USA. · Department of Physics, Faculty of Science, Cairo University, Giza, 12613, Egypt.
Do We Really Need Quantum Machine Learning?: A Multidimensional Empirical Study
May 27, 2026Sudip Vhaduri, Ryan Gammon, Sayanton DibboQuantum Convolutional Neural Networks Quantum Machine Learning
Department of Computer Science, University of Alabama, AL 35487 · Clemson University

Jul 9, 2026J/K move · Enter open · S save

A Novel Parallel QCNN Architecture with Efficient Classical Simulability

Lawrence Nguyen, Hiu Yung Wong

Department of Electrical Engineering San Jos´e State University San Jos´e, USA

Quantum Convolutional Neural Networks Quantum Circuit Simulation Quantum Neural Networks Image Classification

Open paper AskPDF arXiv

This work presents a study of an implementation of a novel Quantum Convolutional Neural Network (QCNN) for binary classification of images from the Modified National Institute of Standards and Technology (MNIST) dataset. Using a novel architecture inspired by previous QCNN and classical convolutional neural network (CNN) implementations, we use a hierarchical partitioning approach to implement a QCNN circuit that can be approximated and simulated efficiently on a classical machine for a large problem. First, the original image is partitioned such that each process handles a smaller portion of the image, which is encoded into independent states. Then, these partitions merge and combine, resulting in states that contain information from both partitions while halving the number of processes. After repeating this until one process remains, we reduce the dimensionality of the state until a single qubit remains for measurement. Using this approach, we can use multiple processes in parallel to simulate a large QCNN program without the need for exponentially growing hardware requirements as the number of qubits increases. In our work, we use this scheme to train a 128-qubit model, which is impossible to run on any classical supercomputer without the novel architecture. We also explore the impact of this new model architecture on prediction accuracy by training it to perform binary classification on the MNIST dataset with a small number of qubits, and comparing it to a model without partitioning. Our initial findings show that partitioning images into smaller sub-images with this architecture does not degrade the model's performance and sometimes even improves it, likely because it reduces the Barren plateaus issue in the partitioning process.

Abstract

Explore similar work

A Novel Parallel QCNN Architecture with Efficient Classical Simulability

Layer-wise QUBO-Based Training of CNN Classifiers for Quantum Annealing

Do We Really Need Quantum Machine Learning?: A Multidimensional Empirical Study