Previously we have discussed using various distance-based techniques for clustering and classifying features obtained from sensor outputs. These techniques are based on looking at the mean and covariance of the features to determine if they belong in a given class. In this section of the course, we will examine the notions of crisp and fuzzy classes. We will also examine some relatively simple techniques for determining the non-linear transfer functions that map features into classes. These transfer functions enable the HCI designer to develop systems that can find patterns in the data obtained from input sensors and map these patterns to specific actions which can control the operation of the computer.
For example, as was discussed in the first homework assignment, we might want to recognize whether someone is nodding their head to mean "yes" or shaking their head to mean "no" when interacting with a computer. Two accelerometers can be used to obtain the data on head movement. However, it is up to the computer to find patterns in the accelerometer data which correspond to the two head gestures. This is a relatively straight forward task as there would only need to be two features (the output level of each accelerometer) and two classes (the two head gestures). The transfer function between features and class would probably be quite obvious. If the forward facing accelerometer's output is high and the side facing accelerometer's output is low, then the individual is nodding. The opposite sensor outputs would indicate a shake..
What if, on the other hand, one wanted to use sign language to communicate with the computer. Many sensor outputs which measured finger, hand, and arm positions would be needed. All or most of these outputs would be used to calculate important features of the gesture. The computer would then have to use some or all of these features to determine which gesture was made. This is an extremely complicated problem with the possibility that an individual might not use exactly the same gesture for a given sign every time. It is also quite likely that each individual might make slightly different gestures for the same sign. The transfer function between sensor features and a sign may be quite different and might have to be changed for each individual.
Pattern recognition using fuzzy sets, which is discussed in this section, is a technique for determining such transfer functions. A good overview of the material discussed is "A Review of Probabilistic, Fuzzy, and Neural Models for Pattern Recognition" by James C. Bezdek in the Journal of Intelligent and Fuzzy Systems, vol.1 (1), pgs. 1-25, 1993.