Training of a neural network with using deterministic transforms

Abstract

Deep neural networks have been a leading research topic within the machine learning field for the past few years. The introduction of graphical processing units (GPUs) and hardware ad- vances made possible the training of deep neural networks. Previously the training procedure was impossible due to the huge amount of training samples required. The new trained introduced architectures have outperformed the classical methods in different classification and regression problems. With the introduction of 5G technology, related to low-latency and online applica- tions, the research on decreasing the computational cost of deep learning architectures while maintaining state-of-art performance has gained huge interest. This thesis focuses on the use of Self Size-estimating Feedforward Network (SSFN), a feed- forward multilayer network. SSFN presents low complexity on the training procedure due to a random matrix instance used in its weights. Its weight matrices are trained using a layer-wise convex optimization approach (a supervised training) combined with a random matrix instance (an unsupervised training). The use of deterministic transforms is explored to replace random matrix instances on the SSFN weight matrices. The use of deterministic transforms automat- ically reduces the computational complexity, as its structure allows to compute them by fast algorithms. Several deterministic transforms such as discrete cosine transform, Hadamard trans- form and wavelet transform, among others, are investigated. To this end, two methods based on features statistical parameters are developed. The proposed methods are implemented on each layer to decide the deterministic transform to use. The effectiveness of the proposed approach is illustrated by SSFN for object classification tasks using several benchmark datasets. The results show a proper performance, similar to the original SSFN, and also consistency across the different datasets. Therefore, the possibility of introducing deterministic transformations in machine learning research is demonstrated.