Integrative Graduate
Education and
Research Traineeship

Highlight

Cost-sensitive learning

pedestrian detections

Projects:

Vision and Learning in Humans and Machines

Institution:

University of California at San Diego

Primary Strategic Goal:

Discovery: Foster research that will advance the frontiers of knowledge, emphasizing areas of greatest opportunity and potential benefit and establishing the Nation as a global leader in fundamental transformational science and engineering.

Secondary Strategic Goal:

Learning: Build the Nation's research capability through critical investments in advanced instrumentation, facilities, cyber infrastructure, and experimental tools.

Achievement/Results

NSF funded researchers Hamed Masnadi-Shirazi and Nuno Vasconcelos at the University of California San Diego, have established a general framework for the design of cost sensitive loss functions that can be used for the derivation of cost sensitive classification algorithms. Affective cost sensitive boosting and support vector machine (SVM) classification algorithms are derived and tested on a variety of cost sensitive tasks such as medical diagnosis and credit fraud detection in machine learning and object detection in computer vision with state of the art results. The figures show two example images for the task of pedestrian and car detection. The design of optimal classifiers with respect to losses that weigh certain types of errors more heavily than others is denoted as cost-sensitive learning . In these problems the cost of missing a target is much higher than that of a false-alarm, and classifiers that are optimal under classic symmetric costs tend to under perform. The principles of cost sensitive classification and loss function design were previously published by Hamed Masnadi-Shirazi and Nuno Vasconcelos in 2007 ICML and 2008 NIPS papers respectively. These principles are combined to establish a series of conditions that provide a generic procedure for the design of cost-sensitive classification algorithms. Such cost sensitive algorithms are guaranteed to satisfy a series of Bayes consistent optimality conditions. Such optimality guarantees were lacking in previous cost sensitive algorithms resulting in inferior performance. This serves as another example where the insight gained from the area of machine learning and statistics has resulted in improved performance in computer vision applications. The cost sensitive boosting algorithms with applications to face detection and car detection have been accepted for publication in IEEE-PAMI 2009 and the cost sensitive SVM algorithm with applications to medical diagnosis and fraud detection have been accepted for publication in ICML 2010.

Address Goals

Improving algorithms for cost-sensitive applications is important for a wide variety of computer vision and machine learning applications for real-world important problems. This research is part of the dissertation work of Hamed Masnadi-Shirazi.