Training Methodology for a Multiplication Free Implementable Operator Based Neural Networks

Yıldız, Ozan
Technological advances opened new possibilities for computing environments including smart phones, smart appliances, and drones. Engineers try to make these devices smart, self-sustaining through usage of machine learning techniques. However, most of the mobile environments have limited resources like memory, computing power and battery, and consequently traditional machine learning algorithms which require relatively high resources might not be suitable for them. Therefore, efficient versions of traditional machine learning algorithms receives interest for these kinds of environments. Recently, an operator named the ef-operator, which avoids multiplication is proposed as an alternative to classic vector multiplication. Recent studies, showed that ef-operator can be used on machine learning problems with small degradation on performance to gain energy efficiency. This thesis concerns with the application of this ef-operator over artificial neural networks. An artificial neural network architecture based of this ef-operator proposed which can approximate any Lebesgue integrable function. Applicability of standard backpropagation algorithm for this new network architecture is analyzed and a modified version of backpropagation algorithm with a line search step proposed for training this network architecture. 
Citation Formats
O. Yıldız, “Training Methodology for a Multiplication Free Implementable Operator Based Neural Networks,” M.S. - Master of Science, Middle East Technical University, 2017.