Speaker
Description
This work proposes and evaluates two novel Quantum Implicit Neural Representation (QINR) models for natural image representation. The first, is a near-fully quantum model adopting a conventional Variational Quantum Algorithm (VQA) framework with minimal classical components, tested with four encoding strategies: standard angle encoding via Pauli rotation gates $\hat{R}_x, \hat{R}_y,$ or $\hat{R}_z$ and Chebyshev feature maps. The second is a novel extension of the state-of-the-art Quantum Implicit Representation Network (QIREN) framework incorporating Chebyshev encoding in place of the original $\hat{R}_z$ layers, evaluated across four different variants: Arctan, Sigmoid, Softsign, and Tanh.
Chebyshev feature maps, which encode data into a Chebyshev polynomial basis through $\hat{R}_y$ rotation gates, have previously demonstrated strong performance in solving nonlinear differential equations with differentiable quantum circuits, but their application within QINRs has not previously been explored. Both architectures use tested Quantum Machine Learning (QML) technique, specifically data reuploading, interleaving Chebyshev encoding layers with trainable Hardware-Efficient Ansatz (HEA) blocks to enhance expressivity within NISQ-compatible circuit depths.
Preliminary results presented in this work demonstrate that Chebyshev encoding improves expressivity across both frameworks relative to standard encoding techniques. These findings suggest that expressive feature map encoding is a promising design principle for QINRs, motivating future investigation into broader classes of structured encodings to achieve near-fully quantum architectures.
| Topic | Quantum variational circuits and their applications |
|---|---|
| Additional Topic | Encoding and processing of data in quantum systems |