/ Qiangfu Zhao / Associate Professor
/ Jintae Lee / Assistant Professor
/ Jie Huang / Assistant Professor
/ Ruck Thawonmas / Assistant Professor
The multimedia devices presently available add sound, still pictures, and moving pictures to traditional computer displays which show only text and diagrams. The addition of these few functions, however, offers remarkable convenience to the user of such multimedia devices. For this reason, it is widely believed that the impact of these devices on human society in the near future will be extremely great.
The properties desirable for multimedia devices can be visualized on a graph with two axes. The horizontal axis represents a device's capacity for depth and variety of communication and the vertical axis represents the capacity for naturalness and high fidelity. Ideally, multimedia devices should be far along both axes if they are to offer the most comfortable and efficient human interfaces. Unfortunately, progress along these dimensions has not gone far enough. Prolonged use of computers still fatigues the body and fails to match the depth and quality of communication that can exist between humans in face-to-face encounters.
For example, human beings can communicate with great ease and variety through the spoken word. Computers can't. In spite of the fact that speech synthesis by computers is almost perfect, speech recognition has not reached the level of practical application that natural speech can achieve. Furthermore speech recognition algorithms are overly sensitive to existing noise. The area of speech recognition is still in the primitive stages and needs to evolve further to make multimedia devices more effective.
Future multimedia systems should be able to accept and recognize natural information. While traditional multimedia systems provide more and more information, future systems should be able to select, and provide useful and important information. One project in this laboratory is to study how to build such kind of intelligent multimedia systems. Our short-term goal is to build a system that can recognize 3D characters and documents. The long-term goal is to construct a system that can recognize nature images and speeches.
Sign language is another means of human communication that computers cannot participate in as easily as those who are forced to use it because of hearing impairment. If a computer could be made to communicate via this medium, far more people could access intellectual services and actively participate in multimedia that have been unable to do so thus far. At present, pioneer research to incorporate this dimension of communication as far as synthesis and recognition technology has just begun in our laboratory.
As for the dimension of high fidelity, the digitalization of high definition TV is one example of progress that has been made along the vertical axis. Signal processing technology is indispensable for the generation, processing, and recognition of visual, auditory and control signals. Furthermore, multimedia devices are preferable to work in real-time. These requirements can be fulfilled not only by high computing processor speeds and large memories, but also by fast algorithms. Therefore, the study of signal processing is fundamental in our laboratory.
Refereed Journal Papers
Compared with the conventional approaches, the evolutionary algorithms (EAs) are more efficient for system design in the sense that EAs can provide higher opportunity for obtaining the global optimal solution. However, in most existing EAs, an individual corresponds directly to a possible solution, and a large amount of computations are required for designing large-scaled systems. To solve this problem, this paper proposes a co-evolutionary algorithm (CEA). The basic idea is divide and conquer: divide the system into many small homogeneous modules, define an individual as a module, find many good individuals using existing EAs, and put them together again to form the whole system. To make the study more concrete, we focus the discussion on the evolutionary learning of neural networks for pattern recognition. Experimental results are provided to show the procedure and the performance of the CEA.
To design the nearest neighbor based multilayer perceptron (NN-MLP) efficiently, we have proposed a non-genetic evolutionary algorithm called the R4-rule. For off-line learning, the R4-rule can produce the smallest or nearly smallest networks with high generalization ability by iteratively performing four basic operations: recognition, remembrance, reduction and review. For on-line learning, we have introduced two parameters: the review limitation beta and the reduction rate gamma. These two parameters are very useful for stabilizing the algorithm, provided that their values are chosen properly. In this paper, we propose a method for finding good parameters based on a simple genetic algorithm (SGA). The efficiency of combining the genetic algorithm with the non-genetic R4-rule is illustrated using experimental results.
Compared with the conventional algorithms, the evolutionary algorithms (EAs) are usually more efficient for system design because they can provide higher opportunity for obtaining the global optimal solution. However, the EAs cannot be used directly to design large-scale systems because a large amount of computations are required. To solve this problem, many approaches have been proposed in the literature. The cooperative co-evolutionary algorithms (CCEA) is possibly one of the most efficient approaches. The basic idea of most CCEAs is divide-and-conquer: divide the system into many modules, define an individual as a candidate of a module, assign a population to each module, find good individuals within each population, and put them together again to form the whole system. In this paper, we generalize our earlier studies, and introduce a society model for the study of CCEAs. Based on the society model, we will formulate existing CCEAs in a general framework. We will also provide several case studies, all of which are interesting topics for future researches.
It is known that decision trees are very efficient for pattern recognition. If we design a decision tree first, and then map it into a neural network, the network so obtained will have far fewer connections than that designed directly by using conventional (say, BP) learning algorithms. This is because, during the design of a decision tree, useful features can be selected automatically, and then be used in an efficient way. In practice, however, decision trees cannot be used easily because the optimal design of decision trees is an NP-complete problem. In this paper, we apply the genetic programming (GP) to design of decision trees. To make the discussion more concrete, we will focus our discussion on a character recognition problem. The efficiency of the GP is verified through experimental results.