QxBranch has released the findings of a novel quantum computing algorithm with the potential of higher accuracy results for a type of machine learning model called convolutional neural networks (CNNs).
Senior Data Scientist Max Henderson, along with a team of QxBranch developers, investigated a new model called quanvolutional neural networks (QNNs), which extends the capabilities of normal CNNs by adding in a new layer to the training system that is enabled by the unique properties of quantum computers. The article, Quanvolutional Neural Networks: Powering Image Recognition with Quantum Circuits was published on April 9, 2019 on arXiv.
Quanvolutional layers transform classical input data using random quantum circuits, producing nonlinear “quantum” feature maps. These feature maps produced by the quanvolutional layers were used in a larger machine learning stack for an image recognition problem, to evaluate the performance of the addition of the quantum feature maps compared to a standard CNN.
The QxBranch team empirically evaluated the impact of the quanvolutional layers and showed that the addition of a quanvolutional layer resulted in networks that were more accurate than classical models. The properties of QNNs make this algorithm ideal for near-term quantum devices, and future research plans to explore the applicability of such algorithms in search of potential quantum advantages they may provide.
“While not yet showing a clear quantum advantage, we believe exploring variations and improvements in methods like the QNN algorithm will help guide us towards understanding the best machine learning applications in practice on near-term quantum devices” says Henderson.
Click here to read the full paper on arXiv.org