Masters Thesis Research Archive
Using Support Vector Machines, Convolutional Neural Networks and Deep Belief Networks for Partially Occluded Object Recognition
Joseph Lin Chu
A Thesis
In
The Department
Of
Computer Science
Presented in Partial Fulfillment of the Requirements
for the Degree of Master of Computer Science (Computer Science) at
Concordia University
Montreal, Quebec, Canada
March 2014
Abstract
Artificial neural networks have been widely used for machine learning tasks such as object recognition. Recent developments have made use of biologically inspired architectures, such as the Convolutional Neural Network, and the Deep Belief Network. A theoretical method for estimating the optimal number of feature maps for a Convolutional Neural Network maps using the dimensions of the receptive field or convolutional kernel is proposed. Empirical experiments are performed that show that the method works to an extent for extremely small receptive fields, but doesn’t generalize as clearly to all receptive field sizes. We then test the hypothesis that generative models such as the Deep Belief Network should perform better on occluded object recognition tasks than purely discriminative models such as Convolutional Neural Networks. We find that the data does not support this hypothesis when the generative models are run in a partially discriminative manner. We also find that the use of Gaussian visible units in a Deep Belief Network trained on occluded image data allows it to also learn to classify non-occluded images.
– (LINK TO CONCORDIA UNIVERSITY SPECTRUM LIBRARY DOWNLOAD) –
– (LOCAL DOWNLOAD) –
…
Masters Research Paper Archive
Analysis of Feature Maps Selection in Supervised Learning Using Convolutional Neural Networks
Joseph Lin Chu and Adam Krzyzak
Department of Computer Science and Software Engineering
Concordia University, Montreal, Quebec, Canada
Published In: Springer Lecture Notes on Artificial Intelligence
Presented At: 27th Canadian Conference for Artificial Intelligence 2014
Abstract
Artificial neural networks have been widely used for machine learning tasks such as object recognition. Recent developments have made use of biologically inspired architectures, such as the Convolutional Neural Network. The nature of the Convolutional Neural Network is that each convolutional layer of the network contains a certain number of feature maps or kernels. The number of these used has historically been determined on an ad-hoc basis. We propose a theoretical method for determining the optimal number of feature maps using the dimensions of the feature map or convolutional kernel. We find that the empirical data suggests that our theoretical method works for extremely small receptive fields, but doesn’t generalize as clearly to all receptive field sizes. Furthermore, we note that architectures that are pyramidal rather than equally balanced tend to make better use of computational resources.
– (LINK TO SPRINGER DOWNLOAD) –
– (LOCAL DOWNLOAD) –
…
Application of Support Vector Machines, Convolutional Neural Networks and Deep Belief Networks to Recognition of Partially Occluded Objects
Joseph Lin Chu and Adam Krzyzak
Department of Computer Science and Software Engineering
Concordia University, Montreal, Quebec, Canada
Published In: Springer Lecture Notes on Artificial Intelligence
Presented At: 13th International Conference on Artificial Intelligence and Soft Computing 2014
Abstract
Artificial neural networks have been widely used for machine learning tasks such as object recognition. Recent developments have made use of biologically inspired architectures, such as the Convolutional Neural Network, and the Deep Belief Network. We test the hypothesis that generative models such as the Deep Belief Network should perform better on occluded object recognition tasks than purely discriminative models such as Convolutional Neural Networks. We find that the data does not support this hypothesis when the generative models are run in a partially discriminative manner. We also find that the use of Gaussian visible units in a Deep Belief Network trained on occluded image data allows it to also learn to classify non-occluded images.
– (LINK TO SPRINGER DOWNLOAD) –
– (LOCAL DOWNLOAD) –
…
Undergraduate Research Archive
Author’s Notes:
The following is an abstract from an undergraduate honours thesis I wrote when I was studying Cognitive Science at Queen’s University. Since it’s completion I’ve gained a number of additional insights and to a certain extent found answers to some of the questions I originally wrote about, though many of those answers led to more questions as well. For instance, after pursuing the question of convolutions and their relation to neural processing, I learned that much work in signal processing and neuroscience has studied in depth this very phenomenon. That, like many of the ideas I had, were not so original after all, I merely needed to encounter the work done in fields outside of my limited experience in computer science and psychology to discover this.
Nevertheless, this thesis was where the direction of my research first coalesced into something tangible, so I offer it as a reference to those curious about my early motivation and thoughts…
SIMULACRUM OF THE MIND’S EYE:
MODIFYING THE NEOCOGNITRON FOR CONSTRUCTING OBJECT REPRESENTATIONS
by
Joseph L. Chu
COGS499 – Research Report Option
Research Proposal
School of Computing
Queen’s University
2009
Abstract
There have been countless theories throughout history regarding the human mind from a wide range of fields of inquiry. Developments in psychology, neuroscience, and computer science led to the advent of neural networks that were used to model such things as visual perception and memory. At the forefront of this trend, the Neocognitron proved that a biologically inspired architecture could be utilized to recognize hand-written characters. Further advances have led to network models that show promise in the field of object recognition. The question that is proposed to be answered is whether such networks could also categorize and classify objects recognized. If so, the potentiality exists to combine such a network with a semantic memory network to produce a visual semantic memory model capable of object representation. This model is divided into two semi-independent stages, the first of which is object recognition, and the second is semantic representation. Following a biological development chronology, focus is placed on the first stage for further study. A prototype network is proposed and three immediate problems with implementation considered: grayscale conversion, network dimensions, and learning algorithm. After discussing the tentative results of the prototype, further research possibilities are looked at, including possible incorporation of the second stage of object representation.
– (DOWNLOAD FULL PDF) –