# Hierarchical quantum circuit representations for neural architecture search

**Authors:** Matt Lourens, Ilya Sinayskiy, Daniel K. Park, Carsten Blank, Francesco Petruccione  
**arXiv:** [2210.15073](https://arxiv.org/abs/2210.15073)  
**Date:** 2022-10-26  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2210.15073v3/

## Key points

- Machine learning with hierarchical quantum circuits, known as Quantum Convolutional Neural Networks (QCNNs), has shown promise for near-term quantum computing.
- QCNNs eliminate the need for manual feature design and can learn high-level features directly from raw data.
- The authors propose a novel framework that combines techniques from Neural Architecture Search (NAS) with QCNN architectures to address the issue of applying NAS to QCNNs due to the lack of a well-defined search space.
- Using this framework, they generate a family of popular QCNN models resembling reverse binary trees.
- The authors evaluate this family of models on a music genre classification dataset called GTZAN and demonstrate the significance of circuit architecture in achieving accurate classification performance.
- A genetic algorithm is employed to perform Quantum Phase Recognition (QPR) as an example of architecture search using their representation, allowing for improved model performance without increasing complexity and avoiding barren plateaus in cost landscapes.
- An open-source Python package is implemented that enables dynamic creation of QCNNs and facilitates search space design for NAS in QCNNs.
- This work introduces a novel framework that combines NAS techniques with hierarchical quantum circuits to enable efficient architecture search in QCNNs.
- The evaluation on a music genre classification dataset highlights the importance of circuit architecture in achieving high-performance models.
- The implementation of an open-source Python package enhances accessibility and promotes further research in this field.

## Summary

Machine learning with hierarchical quantum circuits, known as Quantum Convolutional Neural Networks (QCNNs), has shown promise for near-term quantum computing. Inspired by the architecture of Convolutional Neural Networks (CNNs), QCNNs eliminate the need for manual feature design and can learn high-level features directly from raw data. To address the issue of applying Neural Architecture Search (NAS) to QCNNs due to the lack of a well-defined search space, the authors propose a novel framework that combines techniques from NAS with QCNN architectures. This framework enables the design of search spaces and facilitates architecture search for QCNNs. Using this approach, they generate a family of popular QCNN models that resemble reverse binary trees. To evaluate the importance of circuit architecture in QCNNs, the authors assess this family of models on a music genre classification dataset called GTZAN. The results demonstrate the significance of circuit architecture in achieving accurate classification performance. Additionally, a genetic algorithm is employed to perform Quantum Phase Recognition (QPR) as an example of architecture search using their representation. This approach allows for improving model performance without increasing complexity and enables navigation through cost landscapes to avoid barren plateaus. Finally, an open-source Python package is implemented which not only enables dynamic creation of QCNNs but also facilitates search space design for NAS in QCNNs. In conclusion, this work introduces a novel framework that combines NAS techniques with hierarchical quantum circuits to enable efficient architecture search in QCNNs. The evaluation on a music genre classification dataset highlights the importance of circuit architecture in achieving high-performance models. The implementation of an open-source Python package further enhances accessibility and promotes further research in this field.