Progressive Classification Using Support Vector Machines
- Sunday, 01 November 2009
An approximate classification is generated rapidly, then iteratively refined over time.
An algorithm for progressive classification of data, analogous to progressive rendering of images, makes it possible to compromise between speed and accuracy. This algorithm uses support vector machines (SVMs) to classify data. An SVM is a machine learning algorithm that builds a mathematical model of the desired classification concept by identifying the critical data points, called support vectors. Coarse approximations to the concept require only a few support vectors, while precise, highly accurate models require far more support vectors. Once the model has been constructed, the SVM can be applied to new observations. The cost of classifying a new observation is proportional to the number of support vectors in the model. When computational resources are limited, an SVM of the appropriate complexity can be produced. However, if the constraints are not known when the model is constructed, or if they can change over time, a method for adaptively responding to the current resource constraints is required. This capability is particularly relevant for spacecraft (or any other real-time systems) that perform onboard data analysis.
The new algorithm enables the fast, interactive application of an SVM classifier to a new set of data. The classification process achieved by this algorithm is characterized as progressive because a coarse approximation to the true classification is generated rapidly and thereafter iteratively refined. The algorithm uses two SVMs: (1) a fast, approximate one and (2) slow, highly accurate one. New data are initially classified by the fast SVM, producing a baseline approximate classification. For each classified data point, the algorithm calculates a confidence index that indicates the likelihood that it was classified correctly in the first pass. Next, the data points are sorted by their confidence indices and progressively reclassified by the slower, more accurate SVM, starting with the items most likely to be incorrectly classified. The user can halt this reclassification process at any point, thereby obtaining the best possible result for a given amount of computation time. Alternatively, the results can be displayed as they are generated, providing the user with real-time feedback about the current accuracy of classification.
Computational savings are realized through the guided application of resources only to those items that are estimated to be misclassified. The coarse approximation may suffice for items that can be classified easily, and more computation can be devoted to ambiguous or difficult cases. Thus, the algorithm enables the user to exert direct, dynamic control over the balance between classification speed and accuracy. When constraints on computation time and other resources preclude a totally accurate classification of all the data, this algorithm provides the best possible approximation to the classification of each item, rather than fully classifying only a fraction of the data set and leaving the rest marked “unknown.”
This work was done by Kiri Wagstaff and Michael Kocurek of Caltech for NASA’s Jet Propulsion Laboratory. For more information, download the Tech nical Support Package (free white paper) at www.techbriefs.com/tsp under the Information Sciences category.
The software used in this innovation is available for commercial licensing. Please contact Karina Edmonds of the California Institute of Technology at (626) 395-2322. Refer to NPO-44089.