Quantum machine learning: tecniche e applicazioni

Over the last thirty years, Machine Learning (\ML) has increasingly caught the attention of the scientific community and, more recently, of the general public. With the increasing amounts of data becoming available, it is widely shared that Machine Learning will become even more pervasive as a necessary ingredient for technological progress. Although the computing power has increased quite fast over a couple of decades and new algorithms have come up continuously, the increment of data is much greater than the growth of computers' performance. Speaking of computational power, Quantum Computing (QC), based on principles of quantum mechanics, such as superposition and entanglement, allows to achieve enormously enhanced performances compared to classical computers. This offer the premise to efficiently solve hard classical problems using quantum systems. Since, on one side, machine learning is under pressure from lack of computing power and, on the other, quantum computing offers strong computational abilities, direct the efforts towards techniques and algorithms able to merge these two worlds seems the way to go: this field of research is called Quantum Machine Learning (\QML). The goal of this research, which is part of an advanced and wide project carried out by the NTT Data quantum research team, is to provide after a proper introduction to both Machine Learning and Quantum Computing, an original contribution in the field of image recognition and image classification, proving how the exploitation of quantum technologies can improve the performance of a classical neural network. In particular, we'll use a state-of-the-art simulated quantum preprocessing algorithm \cite{henderson2019quanvolutional} combined with a Convolutional Neural Network (CNN) \cite{rosdyana2019candlestick} to classify binary candlestick images of stock titles. After reproducing, and in some case improving, the results of Rosdyana et al. work, we'll see how applying a quantum, non trainable layer before the first classical layer of our CNN can significantly improve the accuracy of the algorithm. The purpose of this project, anyway, is not to provide the best market prevision tool, but to show how implementing quantum technologies in hybrid frameworks can provide speed-ups and performance improvements.\\ This research has been completed within a wider project of NTT Data research group. Thanks to Antonio Policicchio and Francesco La Ruffa for the continuous support and Alberto Acuto for the helpful supervision.