Adaptive Event Recognition with the use of Limited Training Data
Abstract: This paper presents a novel event recognition system, which is capable of adapting itself to improve its performance on a small set of training data. The event recognition system is represented by a network of events, related to each other by temporal constraints. This symbolic representat...
Main Authors: | , |
---|---|
Other Authors: | |
Format: | Text |
Language: | English |
Subjects: | |
Online Access: | http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.335.9008 http://users.iit.demokritos.gr/~paliourg/papers/CSC98.pdf |
Summary: | Abstract: This paper presents a novel event recognition system, which is capable of adapting itself to improve its performance on a small set of training data. The event recognition system is represented by a network of events, related to each other by temporal constraints. This symbolic representation is particularly suitable to the treatment of overlapping events, which have been overlooked in most of the work on event recognition. Additionally, a method for refining the temporal parameters of the recognition system is presented here. The method uses a small set of preclassified training examples to improve the performance of the system. The principle of minimal model change is used to overcome the sparseness of the training data. Particular emphasis is given to the issue of multiple positive examples, which is prevalent when allowing overlapping events. The new system has been applied to the thematic analysis of humpback whale songs with encouraging results. |
---|