Course Syllabus -- EEL 4817 (H) Current Topics in Machine Learning II Spring 2008 Textbook: There are no lecture notes or textbook associated with this class. Based on the background material of current research results in Machine Learning that the students have been exposed to in Machine Learning I class (EEL 4818(H), taught in the Fall of 2006), the students will be assigned research projects to work on. The projects will be weekly supervised by the faculty, assigning the project, and on a more frequent basis by their graduate students. All students will report progress of their work monthly. The faculty involved in this teaching effort will be: Georgiopoulos (SEECS), A. Gonzalez (SEECS), Ivan Garibay (SEECS), and A. S. Wu (SEECS). Instructors: Michael Georgiopoulos (michaelg@mail.ucf.edu) A. S. Wu (aswu@cs.ucf.edu) Ivan Garibay (igaribay@mail.ucf.edu) Avelino Gonzalez (gonzalez@pegasus.cc.ucf.edu) Targeted Students: Honors students who have taken EEL 4818 (H) (Machine Learning I) in the Fall 2007 are encouraged to take this class. Honors students who have not taken EEL 4818 (H) in the Fall of 2007 can register for EEL 4817(H) (Machine Learning II) as well, since Machine Learning II is a self-contained class. Students who are not Honors students can also register for this class by e-mailing Dr. Georgiopoulos, and after they are given permission to register by the Honors College. Note that if you are an undergraduate student this is an excellent opportunity to know some of the professors in the SEECS in case you have aspirations to pursue a graduate degree at UCF as an MS or a BS+MS student. Other Prerequisites: EEL3801 or COP3032, or STA3032, and ML-I (EEL 4817(H)). The prerequisite of ML-I (EEL 4818(H)) can be waived by e-mailing Dr. Georgiopoulos and explaining to him your academic background; decisions to waive EEL4818(H) will be made on an individual basis. Goal: The goal of this class is to introduce undergraduate students in the CECS to the current research results in Machine Learning. After having exposed the students in ML-I to current research topics of Machine Learning that are of interest to the faculty teaching this class, the students will be assigned appropriate research projects. The students will work on these projects under faculty supervision and under the faculty’s graduate student supervision. The goal is to be able to publish the research findings in conferences and hopefully journals. Course Outline: There is no course outline associated with this class. Examples of potential projects that could assigned to students are listed below. Note that these are simply examples of projects and they do not necessarily represent the projects that could be assigned to students that will take EEL 4817(H). Sample Project 1: Comparison of Fuzzy ARTMAP and Bayesian Classifiers. Project Description: Fuzzy ARTMAP is a neural network architecture introduced by Carpenter, et al., 1992. Since its inception it has been proven to be one of the premier neural network architectures for classification problems. For example, in a fairly recent publication Joshi, et al. (see Joshi, 1997), compared more than 20 classifiers on 7 different machine learning problems. The conclusion of this study was that Fuzzy Min-Max (see Simpson, 1992) that belongs to the same family of networks as Fuzzy ARTMAP gives the best or the second best classification accuracy over all the other algorithms on these machine learning problems. Some of the advantages that Fuzzy ARTMAP has, compared to other neural network classifiers, is that it learns the required task fast, it has the capability to do on-line learning, and its learning structure allows one to explain the answers that the neural network produces. On the other hand, Bayesian learning algorithms that explicitly calculate the necessary probabilities to build the classifier are amongst the most practical approaches to certain types of learning problems. For example, Michie, et al, (1994) and Kohavi (1995) provided a detailed comparison of the naïve Bayes classifier to other machine learning algorithms including decision trees and neural networks (not Fuzzy ARTMAP). These researchers showed that the naïve Bayes classifier is competitive with these other learning algorithms, and in some cases it outperforms these other methods. Therefore a comparison between these two very competitive classifiers is worth pursuing. The comparison will first focus on artificially generated data. One of the advantages of working with artificial databases first is that we can produce as many data-points as we want. Another advantage of the artificially generated data is that we can experiment with different values for the dimensionality of the input space of the pattern classification task, different values of the number of output categories that the data belong to, and different degrees of overlap amongst the data that belong to different classes. Our belief is that the comparison results will be dependent on all of these variables. Once our comparison is complete for the artificially generated data we will extend the work to real databases. An important part of this work is to meaningfully design the experiments so that we can draw some useful conclusions from this effort. An essential element of this effort is also to establish ways of correctly assessing the performance of the Bayes classifier and the Fuzzy ARTMAP classifier. One measure of comparison is the classification accuracy of the classifiers on a data set that is called test set (different than the training set). Other measures of comparison are the required memory needed to build the Bayes or the Fuzzy ARTMAP classifiers, and the computations needed by each classifier to produce a response to a specific input. A pertinent subject associated with this research effort is how confident we are in the assessments that we make about the goodness of each one of these algorithms. Ways of making confident assessments about the goodness of algorithms is extensively discussed in Dietrrich (1998) and in a Machine Learning book by T. M. Mitchell (see Mitchell, 1997, "Evaluating Hypotheses" chapter 5). Knowledge that the student will be exposed to in this project: The student who will be involved in this project will be exposed to the following papers and book material: Fuzzy ARTMAP (see Carpenter, 1992), Bayes Classifier (see Mitchell, 1997, Bayesian Learning chapter 6), estimating the probabilities associated with the Bayes classifier (see Specht, 1992), evaluating hypotheses (see Mitchell, 1997, Evaluating Hypotheses chapter 5), and others. The faculty and his/her graduate student involved with this project will help the UG student understand this material. The student will also be involved in coding the Bayes classifier. The code of the Fuzzy ARTMAP network will be provided to the student. Expected Results: The results of this effort will be a detailed comparison of 2 powerful classifiers and associated conclusions that will help researchers in the field choose one classifier versus the other. The results of this work will be published in journals and conferences. Sample Project 2: Genetic Algorithm Visualization Project Description: Traditional methods of evaluating genetic algorithm (GA) performance focus on the average time to find a solution or the average quality of solutions found over multiple runs. While such measures can provide estimates of the expected performance of a particular GA configuration, they do not reveal very much information on how a GA found a solution. To fully understand how a GA works, it is important to understand both the expected end and the means to the end. The large numbers of complex, non-linear interactions that compose GAs make them difficult to analyze and a challenge to understand. While all of the data from a GA run can easily be saved into files, accessing and interpreting such a large amount of information is no trivial task. The development of tools for sorting, organizing and displaying such databases could greatly facilitate the access and analysis of such data. Graphical visualization techniques are some of the simplest and at the same time most powerful methods for analyzing and communicating information (Tufte 1983). Well-designed graphical elements can convey large amounts of information in very concise and compact formats. In addition, the human vision system is extremely sensitive to graphical patterns, making graphical representations an extremely useful analysis tool. We are in the process of designing an offline visualization tool to aid in analysis of GA behavior (Wu, et al. 1999). Our system takes advantage of the effectiveness of graphical representations and the flexibility of "clickable'' links to provide a navigation tool for accessing and viewing data from a GA run. Knowledge to which the student will be exposed to in the project: The UG student will work closely with the faculty and with graduate students involved with this project to determine useful methods for displaying GA data from ongoing projects. The UG student will gain an understanding of the basic GA, learn basic computer graphics programming methods, and become familiar with basic scientific data visualization techniques. Expected Results: The UG student will develop visualization techniques and add them to an existing basic visualization program. Techniques will include display of individual solutions as well as data collection/display over a population. The UG student's work on this project will benefit all ongoing projects in the PI's laboratory by providing new methods for evaluating and understanding individual GA runs. In addition, there are plans to publish a paper describing the complete GA visualization system as well as making it available for other researchers in the GA community to use. Meeting Schedule: Each one of the instructors in the class will have the opportunity to present to class a few of the machine learning research projects that he/she is interested in involving the EEL 4817 (H) students with. As of now, the schedule is as follows: Lecture 1: Georgiopoulos (SEECS) Lecture 2: Garibay (SEECS) Lecture 3: Gonzalez (SEECS) Lecture 4: A. S. Wu (SEECS) It is expected that by the third week of January, each one of the students in this class would have picked up a project to work on. After January, each student(s) will interact with the individual professor who is sponsoring the project, and with the professor’s graduate students. After January, at the third week of every month, we will meet as a group (all PIs, all Machine Learning II students, and potentially graduate students) to discuss the progress of the work. We anticipate having 3 group meetings of this nature. At the end of the semester the students will submit a typed report that explains in detail the project that they have worked on, the results that they have produced and discussion of these results. It is anticipated that an effort will be made for all, or most, of the reports to be submitted for publication in conferences and journals. Every professor might have his or her individual preferences on what the report should contain. The professor who sponsors the machine-learning project will also grade the report. An example of the specific contents within a report are: (a) an introduction/problem statement, (b) main body where you address all the tasks (in case the project is broken down in individual tasks), you discuss the neural network classifiers (in your own words) and elaborate on their differences/similarities, (c) Experiments/Results portion (where you discuss the data that you have generated and experimented with and the derived results), and (d) finally a conclusions section where you summarize your conclusions. Create tables and figures wherever appropriate in your report. Your grade will depend on completeness of your work, conciseness and clarity of your work and innovativeness. Course Materials: Course materials will be provided by the supervisor of each project to the student. References: Carpenter, G. A., Grossberg, S., and Reynolds, J. H., “ARTMAP: Supervised real-time learning and classification of non-stationary data by a self-organizing neural network,” Neural Networks, Vol. 4, pp. 565-588, 1991b. Carpenter, G. A., Grossberg, S., and Rosen, D. B., “Fuzzy ART: Fast stable learning and categorization of analog patterns by an adaptive resonance system, ” Neural Networks, Vol. 4, No. 6, pp. 759-771, 1991c. Carpenter, G. A., Grossberg, S., and Reynolds, J.H., "Fuzzy ARTMAP: A neural network architecture for incremental supervised learning of analog multi-dimensional maps," IEEE Transactions on Neural Networks, Vol. 3, No. 5, pp. 698-713, 1992. Dieterich, T. G., "Approximate statistical test for comparing supervised classification learning algorithms," Neural Computation, Vol. 10, No. 7, pp. 1895-1923, 1998. Joshi, J., Ramakrishnan, N., Houstis, E.N., and Rice, J.R., "On neurobiological, neuro-fuzzy, machine learning, and statistical pattern recognition techniques, " IEEE Transactions on Neural Networks, Vol. 8, No. 1, pp. 18-31, 1997. Kohavi, R., Becker B., and Sommerfield, D., "Improving the Simple Bayes," The 9th European Conference on Machine Learning, 1997. Michie, D., Spiegelhalter, D. G., and Taylor, C.C., Machine Learning, neural and statistical classification, (edited collection). New York: Ellis Horwood, 1994. Mitchell, T., M., Machine Learning, McGraw-Hill Companies, Inc., 1997, New York. Simpson, P. K. "Fuzzy min-max neural networks -- Part 1: Classification", IEEE Transactions on Neural Networks, Vol. 3, No. 5, pp. 776-786, Sept. 1992. Tufte E. R. The Visual Display of Quantitative Information, Graphics Press, 1983. Wu, A. S., De Jong, K. A., Burke, D. S., Grefenstette, J.J., and Ramsey, C.L., "Visual analysis of evolutionary algorithms", In Proceedings of the 1999 Congress on Evolutionary Computation, 1999. Grading: Project Assignment: 100% (The grading of the project will be based on the frequent interactions that the faculty and the students have during the project completion process and a detailed typed project that the students have to submit at the end of the class). The letter grade policy is as follows: A: 90 and above B: 80-89 C: 70-79 D: 60-69 F: < 60 Note that the aforementioned guidelines are approximate and it is up to the instructors’ discretion to change them, provided that prior notice is given to you.