Adaptive and Hybrid Machine Learning Approaches Utilizing General Purpose Computing on Graphical Processing Units

Unlike machines, humans and animals have very complex reasoning capability that allows them to adapt to changes in the naturally world, while computers tend to be very limited in that same aspect. What limits machines from becoming adaptable can span many topics, but of these attributes which would limit a machine's ability to adapt is its inability to process input not specifically designed for it. The purposes of this study is to view and work with some possible solutions which may assist with the problems that keep compute devices rather inflexible. The focus of this study will span across two particular domains. The first is the hardware architecture which determines how tasks are executed on a given machine. The second is the usage of machine learning approaches, which are classes of algorithms that rely on the nature of the input rather than a programmer's explicit instructions for decision making. The nature of the data evaluated in this work is sequential as a great variety of information found in the natural world are not discrete. This work will evaluate and contrast how one is to determine for which case would a particular architecture or a learning approach should be used to maximize the expected performance and accuracy of a system. Emphasis will be placed on the usage of Heterogeneous architecture and the usage of algorithm that is either a hybrid of multiple algorithms or adaptable in nature so that it may account for unexpected variation in the data that is likely to vary over time.