Speaker (4:00 - 4:30)
Discriminative Density Ratio for Classification Domain Adaptation
Domain adaptation deals with the challenging problem that has arisen from the distribution difference between the training and the test data. An effective approach is to reweight the training samples to minimize that discrepancy. To be specific, many methods are based on developing Density-ratio (DR) estimation techniques and then applying cost-sensitive learning algorithms to the reweighted data. Although these methods work well for regression problems, their performance on classification problems is unsatisfactory. This is due to the fact that these methods focus on matching the sample marginal distributions without paying attention to preserving the separation between classes in the reweighted space. In this paper, we propose a novel method for Discriminative Density-ratio (DDR) estimation that addresses the aforementioned problem by estimating the density ratio of joint distributions in a class-wise manner. Our approach combines three learning objectives: minimizing generalized risk on the reweighted training data, minimizing class-wise distribution discrepancy, and maximizing the separation margin on the test data. To solve the complex DDR problem, two implementations are presented on the basis of block-coordinate update method. We provide analysis on the theoretical convergence of the presented implementations and demonstrate their effectiveness with extensive experiments conducted on sampling bias scenarios and cross-dataset tasks.
Speaker (4:30 - 5:00)
Samaneh Hosseini Semnani
Multi-Target Tracking in Large-Scale Surveillance Systems
Sensor Networks is an emerging field of study that is expected to touch many aspects of our life. Sensor Networks are well studied and efficiently used in mobile surveillance applications. One key challenge in large-scale surveillance systems is the mobility control and coordination of sensors. One of the most powerful algorithms that has attracted many researchers in the field of mobility control in recent years is the Flocking algorithm . Flocking is a form of cooperative behavior of large number of autonomous interacting agents able to show coordinated group behavior. The self-organizing feature and local communication requirement of flocks can provide deeper insight into the management of sensors in sensor networks.
This work discusses two flocking-based mobility control models, Flocking and Anti- Flocking, used to organize multi-mobile sensor networks in a surveillance area. Although each of these approaches has positive characteristics, each also has its special drawbacks in a surveillance application. Flocking algorithm focuses only on target coverage, while Anti-Flocking algorithm focuses only on area coverage. Semi-Flocking algorithm is presented in this work to fill this gap. The main idea of Semi-Flocking algorithm is to create small flocks of sensors around each target while still leaving some sensors free to search the environment for detection of new targets. The proposed Semi-Flocking algorithm exhibits outstanding performance in meeting the objectives of surveillance application.