Learning from Small Samples with High Dimensionality
Project Award Date: 09-01-2006
Multi/hyperspectral sensors and multidimensional data acquisition systems are becoming increasingly prevalent in various applications. Target detection/recognition and other operations require the classification of small samples with high dimensionality. However, most current classification methods assume the availability of large sample sizes.
ITTC researchers are developing innovative approaches of feature selection and classification for small samples with high dimensionality. A structural risk minimization (SRM)-based selection method will collect a subset of features that are guaranteed to generalize well for small sample problems. Investigators are creating practical methods for optimizing kernels and selecting the parameters of kernels in support of vector machine classifiers.
This research will lead to a significant improvement in classifying small samples. Furthermore, as multidimensional data acquisition systems and new sensor technologies are increasingly used in military and civil applications, the proposed research will provide new capability to fully and efficiently exploring the data acquired from these sensors.