Practical insights on the effect of different encodings, ansätze and measurements in quantum and hybrid convolutional neural networks

Jesús Lozano-Cruz ‡, Albert Nieto-Morales ‡, Oriol Balló-Gimbernat, Adan Garriga, Antón Rodríguez-Otero, Alejandro Borrallo-Rentero · 2025

Read Practical insights on the effect of different encodings, ansätze and measurements in quantum and hybrid convolutional neural networks by Jesús Lozano-Cruz ‡, Albert Nieto-Morales ‡, Oriol Balló-Gimbernat, Adan Garriga, Antón Rodríguez-Otero, Alejandro Borrallo-Rentero (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-20355.txt.

Introduction

Jesus Lozano-Cruz and co-authors test, in a 2025 study, which design picks matter most in quantum neural networks for image work. Their test bed sorts satellite images from the EuroSAT set. They built about 500 model forms, split between pure quantum and hybrid forms that mix quantum and classic parts. Each form varied in how data is coded in, the shape of its quantum circuit, and how the result is read out. Hybrid forms were matched against the same form with the quantum part cut out.

For hybrid forms, data coding ruled: scores swung by more than 30 percent across coding picks. Circuit shape and read-out picks moved scores by less than 5 percent. For pure quantum forms, the read-out plan mattered most, with a swing near 30 percent. The ranks give builders a clear place to spend effort first.

Catalog record

CREATORJesús Lozano-Cruz ‡, Albert Nieto-Morales ‡, Oriol Balló-Gimbernat, Adan Garriga, Antón Rodríguez-Otero, Alejandro Borrallo-Rentero
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE50 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-20355

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Practical insights on the effect of different encodings, ansätze and measurements in quantum and hybrid convolutional neural networks Jesús Lozano-Cruz ‡, Albert Nieto-Morales ‡, Oriol Balló-Gimbernat, Adan Garriga, Antón Rodríguez-Otero, Alejandro Borrallo-Rentero Abstract This study investigates the design choices of parameterized quantum circuits (PQCs) within quantum and hybrid convolutional neural network (HQNN and QCNN) architectures, applied to the task of satellite image classification using the EuroSAT dataset. We systematically evaluate the performance implications of data encoding techniques, variational ansätze, and measurement in ≈ 500 \approx 500 distinct model configurations. Our analysis reveals a clear hierarchy of influence on model performance. For hybrid architectures, which were benchmarked against their direct classical equivalents (e.g. the same architecture with the PQCs removed), the data encoding strategy is the dominant factor, with validation accuracy varying over 30 % 30\% for distinct embeddings. In contrast, the selection of variational ansätze and measurement basis had a comparatively marginal effect, with validation accuracy variations remaining below 5 % 5\% . For purely quantum models, restricted to amplitude encoding, performance was most dependent on the measurement protocol and the data-to-amplitude mapping. The measurement strategy varied the validation accuracy by up to 30 % 30\% and the encoding mapping by around 8 8 percentage points. Keywords: quantum machine learning, parameterized quantum circuit, image classification, quantum neural network, ansatz ‡ ‡ footnotetext: Corresponding authors: [email protected] and [email protected] In recent years, quantum machine learning (QML) has emerged as an interdisciplinary field at the intersection of quantum computing and machine learning Schuld et al. 2014a ; Wittek 2014 ; Biamonte et al. 2017 ; Schuld and Petruccione 2018 . Its objective is to leverage phenomena in quantum mechanics to tackle problems from machine learning. A central component of many contemporary QML models is the parameterized quantum circuit (PQC), a flexible framework for creating trainable quantum circuits that can be used independently or integrated as hybrid components within classical neural networks (NNs). This work investigates the use PQCs for the task of satellite image classification using the EuroSAT dataset Helber et al. 2019 . These PQCs serve as trainable quantum components that can be integrated into NNs or used by themselves. While PQCs offer a flexible paradigm for hybrid and purely QML models, their efficacy is highly sensitive to a multitude of design choices, with the most important being: the encoding technique, the ansätze and the measurement strategy. We study two different hybrid architectures networks and a purely quantum one, all inspired by convolutional neural networks (CNN). For the hybrid ones we have a “quantum first” model, in which PQCs are used as convolutions to extract features, and a “quantum later” model Senokosov et al. 2024 , in which PQCs are used between the fully connected layers to process the extracted features Senokosov et al. 2024 ; Hur et al. 2022 ; Elhag et al. 2024 . For the quantum architecture, we use three variations of PQCs as standalone components, one of which inspired by the QCNN, introduced in Cong et al. 2019 . Our approach builds on the work of Bowles et al. 2024 , which systematically benchmarked different quantum machine learning models by varying the encoding, measurement and hyperparameters. Our difference lies in the fact that we fix the hyperparameters and benchmark our hybrid and purely quantum models across all three key circuit components: encoding, ansätze and measurement. The primary objective of this work is not necessarily to achieve state-of-the-art performance, or claim quantum utility, but rather to find practical insights into how core PQC design elements influence model behavior and efficacy. The findings are intended to aid researchers and practitioners in designing QML models. All our experiments were done through simulations using the CTIC Quantum Testbed (QUTE) platform García Mangas and Alonso 2025 . We used Pennylane Bergholm et al. 2022 to implement the PQCs, both in the purely quantum and hybrid models, and PyTorch P. et al. 2019 for the classical layers. .1 Image classification Our focus will be in classifying satellite images from the EuroSAT dataset Helber et al. 2019 , examples can be seen in FIG. 1 . The dataset consists of 27,000 geo-referenced images, provided by the Copernicus Earth observation program European Space Agency nd using the Sentinel-2 satellite, and comes in two versions: RGB and all bands. The first consists of only RGB channels and the other uses all the bands of the spectrum as collected from Sentinel-2. Each image channel has a resolution of 64\times 64 pixels. In our experiments, we used the RGB version and downscaled them to 32\times 32 and 16\times 16 . The former was only studied for the hybrid models, due computation constraints, and the latter for both hybrid and purely quantum models. We selected this dataset as a continuation of the work conducted within the ARQA Cervera Technology Transfer Network Red ARQA 2024 . Nevertheless, identical experiments can be replicated on alternative datasets, provided they satisfy the previously established size constraints. Figure 1 : Samples from the EuroSAT dataset introduced in Helber et al. 2019 . Each sample is a 64\times 64\times 3 RGB image obtained from the Sentinel-2 satellite. .2 Parameterized quantum circuits PQCs are the cornerstone of our approach. We integrate them as layers consisting of one or more circuits in the hybrid models and use them by themselves in the purely quantum ones. In general, PQCs consist of three key components: encoding, used to embed the information in a quantum state, variational ansatz, used to manipulate the encoded information, and measurement, used to extract information from the quantum state. FIG. 2 shows the prototypical example of a PQC. {Embedding} {Embedding} {Ansatz} {Ansatz} A A A A Classic input Classic output Figure 2 : Prototypical quantum circuit scheme. .2.1 Data encoding As previously mentioned, the first stage in a PQC is the data encoding, this step serves as a mapping from data to a quantum state. Given that there is not a one size fits all solution for the encoding step, we chose to include the following ones into our analysis: • Angle encoding Schuld and Petruccione 2018 ; Schuld et al. 2014b ; Cao et al. 2017 : This is a prevalent and relatively straightforward method where N classical data features, x_{1},x_{2},\ldots,x_{N} , are encoded into the rotation angles of N individual qubits. Each feature x_{j} is used as an angle parameter for a single-qubit rotation gate, such as R_{X}(\theta_{j}) , R_{Y}(\theta_{j}) , or R_{Z}(\theta_{j}) , applied to a specific qubit. Notably, using R_{Z}(\theta_{j}) as an encoding gate acting on an initial initial state from the computational basis only adds an irrelevant global phase. • Amplitude encoding Schuld and Petruccione 2018 : This technique embeds a classical data vector of 2^{n} normalized real numbers, \{\alpha_{i}\} , directly into the amplitudes of an n -qubit quantum state vector |\Psi\rangle . The state is represented as: |\Psi\rangle=\sum_{i=0}^{2^{n}-1}\alpha_{i}|i\rangle where |i\rangle are the computational basis states, and the amplitudes must satisfy the normalization condition \sum_{i=0}^{2^{n}-1}|\alpha_{i}|^{2}=1 . The state preparation process required to achieve an arbitrary state |\Psi\rangle can be computationally expensive, potentially demanding an exponentially deep circuit Knill 1995 . • Instantaneous quantum polynomial-time (IQP) encoding Havlíček et al. 2019 : This encoding extends the concept of angle encoding by

Read the complete text