Bilkent
University
Department of Computer Engineering
S E M I N A R
Indian
Statistical Institute
Literature on feature selection deals with methods that are "off-line" in nature. Typically, some index is computed on each feature or on subsets of features. Then features are either ranked or some subset is selected. Either a few top ranked features or the selected subset is then used to design the system. Ranking is not a good philosophy, because the two top-ranked features are not necessarily the best two (eg., correlated features). It is worth noting here that quality of a feature depends on the TOOL being used and the PROBLEM being solved. So we introduce a novel concept of "ON-LINE" feature selection where the system picks up the required features along with training of the system - the idea is to associate a gate with each feature and keep the gate almost closed at the beginning of the training. The training process opens these gates (features), which are important and shuts the gates more tightly which are bad or redundant. In this context, we will explain three systems. The first system is designed for the multi-layer perceptron type networks. The system is applicable to both classification and function approximation type problems. The second system is built based a neuro-fuzzy framework, which uses a five layer network for solving function approximation type problems. Finally, this system is modified for dealing with classification type problems. The second part deals with sensor selection or group feature selection. (to my knowledge this is also for the first time) For many applications, the input comes from different sensors. For example, in case of an intelligent weld inspection system, the sensors could be X-ray image, Acoustic emission, eddy current and so on. The signal obtained from each sensor is used to compute several features. For example, the X-ray image can be used to compute several co-occurrence based features. In such cases, a more challenging problem comes - selection of sensors (in other words, selection of groups of features, where each group is computed using the signal obtained from a particular sensor). Clearly, if the number of necessary sensors can be reduced, the hardware cost of the system, the design complexity of the system and the cost (both in terms of time and money) of decision making can be drastically reduced. This problem, to our knowledge, has not been addressed in the literature. We will discuss two systems for online sensor selection. The first approach is applicable to multi-layer perceptron type networks while the second method is for radial basis function type network.
DATE: June 30, 2003, Monday @ 14:30
PLACE: EA-409