Date of Award

Winter 2018

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Electrical/Computer Engineering

Committee Director

Khan M. Iftekharuddin

Committee Member

Chunsheng Xin

Committee Member

Norou Diawara

Committee Member

Hongyi Wu

Abstract

Recognition is fundamental yet open and challenging problem in computer vision. Recognition involves the detection and interpretation of complex shapes of objects or persons from previous encounters or knowledge. Biological systems are considered as the most powerful, robust and generalized recognition models. The recent success of learning based mathematical models known as artificial neural networks, especially deep neural networks, have propelled researchers to utilize such architectures for developing bio-inspired computational recognition models. However, the computational complexity of these models increases proportionally to the challenges posed by the recognition problem, and more importantly, these models require a large amount of data for successful learning. Additionally, the feedforward-based hierarchical models do not exploit another important biological learning paradigm, known as recurrency, which ubiquitously exists in the biological visual system and has been shown to be quite crucial for recognition.

Consequently, this work aims to develop novel biologically relevant deep recurrent learning models for robust recognition using limited training data. First, we design an efficient deep simultaneous recurrent network (DSRN) architecture for solving several challenging image recognition tasks. The use of simultaneous recurrency in the proposed model improves the recognition performance and offers reduced computational complexity compared to the existing hierarchical deep learning models. Moreover, the DSRN architecture inherently learns meaningful representations of data during the training process which is essential to achieve superior recognition performance. However, probabilistic models such as deep generative models are particularly adept at learning representations directly from unlabeled input data. Accordingly, we show the generalization of the proposed deep simultaneous recurrency concept by developing a probabilistic deep simultaneous recurrent belief network (DSRBN) architecture which is more efficient in learning the underlying representation of the data compared to the state-of-the-art generative models. Finally, we propose a deep recurrent learning framework for solving the image recognition task using small data. We incorporate Bayesian statistics to the DSRBN generative model to propose a deep recurrent generative Bayesian model that addresses the challenge of learning from a small amount of data. Our findings suggest that the proposed deep recurrent Bayesian framework demonstrates better image recognition performance compared to the state-of-the-art models in a small data learning scenario. In conclusion, this dissertation proposes novel deep recurrent learning pipelines, which utilize not only limited training data to achieve improved image recognition performance but also require significantly reduced training parameters.

DOI

10.25777/1527-2348

ISBN

9780438991644

Share

COinS