The implementation of the P-N learning process Essay

1. A set of examples X is called an unlabeled set, Y is called a set of labels, and L= {(x,y)} is called a labelled set.
2. The task of P-N learning is to learn a classifier f X ->Y from labelled set Ll and bootstrap its performance by the unlabeled set Xu.
3. Classifier f is a function from a family F parameterized by . The family F is subject to implementation and is considered fixed in training, the training therefore corresponds to estimation of the parameters ..
4. The training process is initialized by inserting the labelled set L to the training set. The training set is then passed to supervised learning, which trains a classifier, i.e., estimates the initial parameters .
5. The learning process then proceeds by iterative bootstrapping. In iteration k, the classifier trained in previous iteration classifies the entire unlabeled set. The classification is analyzed by the P-N experts, which estimate examples that have been classified incorrectly. These examples are added with changed labels (due to experts) to the training set. Each iteration k finishes by retraining the classifier, i.e., estimation of k. The process iterates until convergence or other stopping criterion.
6. The crucial element of P-N learning is the estimation of the classifier errors. The key idea is to separate the estimation of false positives from the estimation of false negatives.
7. P-expert analyzes examples classified as negative, estimates false negatives, and adds them to training set with positive label. In iteration k, P-expert outputs n+(k) positive examples.
8. N-expert analyzes examples classified as positive, estimates false positives, and adds them with negative label to the training set. In iteration k, the N-expert outputs n-(k) negative examples.
9. The P-expert increases the classifier's generality. The N-expert increases the classifier's discriminality.
10. Labelling is plausible since the object appears at one location in each frame and the detected locations build up a trajectory in time. In other words, the labels of the patches are dependent. We refer to such a property as structure. The key idea of the P-N experts is to exploit the structure in data to identify the detector errors.
11. P-expert exploits the temporal structure in the video and assumes that the object moves along a trajectory. If the detector labelled the current location as negative (i.e., made false negative error), the P-expert generates a positive example.
12. The N-expert analyzes all responses of the detector in the current frame and the response produced by the tracker and selects the one that is the most confident.
13. Patches that are not overlapping with the maximally confident patch are labelled as negative. The maximally confident patch reinitializes the location of the tracker.
14. False negatives retrieved by the P-expert are labelled positive and their addition to the set increases detector's generality (it recognizes more appearances of the object), while false positives are labelled as negative by N-expert and increase detectors ability to discriminate against everything that is not an object of interest.