Abstract:
Image classification is a fundamental task in machine learning, with applications in medical
imaging, remote sensing, and autonomous systems. However, the increasing complexity
and high-dimensional nature of image datasets pose significant challenges for conventional
classification models. To address these limitations, this study explores a novel
approach based ondimensionality reduction with variational encoders to enhance feature
representation and improve classification accuracy. Variational autoencoders (VAEs) have
emerged as a powerful tool for reducing dimensionality while preserving essential features.
In this work, we extend VAEs with quantum-enhanced encoding techniques using subsystem
purification, enabling efficient data compression while mitigating information loss.
By leveraging hybrid quantum-classical architectures, we propose a framework that integrates
quantum variational encoders with classical neural networks, optimizing feature
extraction for classification tasks. Our approach is evaluated on benchmark datasets, including
Iris and MNIST, demonstrating improved accuracy and computational efficiency
compared to traditional dimensionality reduction techniques. The results show that our
model achieves superior performance by maintaining discriminative features while reducing
data complexity. Furthermore, we address key challenges such as barren plateaus
and optimization instability through time-nonlocal optimization strategies. This research
contributes to the advancement of quantum-enhanced machine learning by introducing
a robust and scalable dimensionality reduction framework for image classification. The
proposed method sets a new standard for efficient quantum-classical hybrid models, offering
a promising direction for future developments in quantum-assisted data processing
and classification.
