About
Joseph L. Chu is a graduate student in the Master of Computer Science program at Concordia University in Montreal.
He previously studied and completed a Bachelor of Computing with a Subject of Specialization in Cognitive Science at Queen's University in Kingston.
Contact Info:
E-mail: jo_chu@encs.concordia.ca
Teaching Assistantship
I was previously a teaching assistant for a weekly Tutorial for COMP248, during the Fall semester of 2010.
Research
Working Proposal for Thesis
The biological plausibility of connectionism possesses a certain degree of attractive intuitiveness. It just makes a lot of sense that to create artificial intelligence, we should be looking closely at natural intelligences, given that they have been developed and refined by millions of years of natural evolutionary genetic algorithms. A reasonable hypothesis then is that such intelligence is a property of the emergent behaviour of highly structured, hierarchical, and recurrent neural networks.
Consistent with this ground up approach to understanding intelligence is narrowing the focus of research to looking at the inputs of this system, primarily the well-studied visual object recognition system. While others may be content to try to mimick intelligence by using rules to manipulate symbols, I am more interested in understanding the very nature of symbols in the mind, and am pursuaded by the idea that the nature of how perceptions are represented as, and reconstructed in symbols affects how such symbols can be utilized. A symbol's meaning is what it represents. Fundamentally, even an abstract symbol or idea like say 'democracy', is associated with more concrete symbols such as 'people', of which there exists perceivable examples in the real world. 'Democracy' means nothing without the perception of those physical objects that can be connected to the idea. And yet most computers are mere physical symbol systems, manipulating meaningless symbols according to programmed rules. Thus, it make sense to seek to understand the sensory perceptions that give rise to cognition, in order that machines might also be endowed with this fundamental element of intelligence.
To do this I am looking towards the manner in which hierarchical neural network architectures, in particular the convolutional neural networks such as LeCun's LeNet 5, or Fukushima's Neocognitron, are able to perform object and pattern recognition. Clearly in this process, some equivalent or analogue to a learned representation or memory exists.
Historically most convolutional neural networks have been mostly feedforward networks. However, the human brain is rife with neural feedback connections, and the many structural loops this creates are perhaps analoguous to the manner in which thinking often seems to loop about different connections and ideas. Indeed the notion of such 'Strange Loops' being an essential characteristic of consciousness, as popularized by Hofstader, seems to support the idea. Thus, recurrent neural networks are a more realistic analogue, and applying their peculiar properties to the object recognition problem is also a part of my research interests.
As a part of this, I hope to take advantage of the significant strides made in computer vision in understanding how the highly structured hierarchy of simple and complex neuron cells are able to produce what amounts to signal processing through convolutional filtering and feature extraction. Thus, some of my research may involve experimenting to see just how close an analogue certain neural networks, such as Hinton's Deep Belief Networks, are to actual neuromorphic networks.
That is the direction I would like my research to contribute to. Of course, one must always be aware that scientific research could lead to unexpected discoveries, and I am willing to be wrong, if I can be certain of it. The goal of this research is not to confirm what I want to be true, but to test my hypothesis against reality, and in so doing, find genuine knowledge. It is my hope that this knowledge can then serve humanity in some way, perhaps in developing more intelligent, more aware machines that can be utilized for more economic and scientific endeavours. Or perhaps a better understanding of the human mind through its accurate simulation in machinery will be sufficient reward for these efforts.
Primary Research Interests:
- Neural Networks and Object Recognition
- Convolutional Neural Networks
- Recurrent Neural Networks
- Unsupervised Learning
- Pattern Recognition
- Computer Vision
- Object Knowledge Representation
- Visual Semantic Memory
Secondary Research Interests:
- Cognitive Psychology Models
- Holographic Memory
- Connectionism
- Artificial Intelligence
- Human-Computer Interaction
- Social Psychology of Online Games
- Behavioural Economics
- Philosophical and Political Implications of Cognitive Science
Last Updated: March 10th, 2010