Browsing by Subject "Restricted Boltzmann Machines"

Quantum Computing is a novel technology that has wide applicability in the field of machine learning. One of these applications is training Quantum Restricted Boltzmann Machines, which have been shown to have advantages over their classical counterparts. These Quantum Restricted Boltzmann Machines can be then used to pretrain more complex machine learning models, such as Deep Belief Networks, which means that quantum annealing can have applications in the field of deep learning. Main issue of Quantum Restricted Boltzmann Machines is that embedding them into quantum annealing devices will restrict their layer size and connectivity quite drastically. This thesis proposes the use of a common weight regularization method called the unit dropout method to reduce the overall size of these networks by splitting these Restricted Boltzmann Machines into smaller more manageable models, training them separately and composing them into a complete model. While this method can be shown to affect learning negatively, it is yet to be known, whether the advantages of quantum computing can outweigh the disadvantages of extreme use of the unit dropout method.