Audio perception and information retrieval and Interactive Intelligent Computing

This will be a short overview of audio content-based retrieval, and applications in film soundtrack analysis and psychoacoustics that are in the planning stages. In this talk, I will introduce a concept of "interactive intelligent computing" which combines the statistical computing theories with the HCI-type of human intervention. In past decades, we have seen burgeoning of many interesting computing paradigms such as pattern recognition, machine learning, and data mining to name a few. What they all share is their main goal to 1) learn statistical regularity from given data and 2) realize automation of certain computing tasks, such as segmentation, classification, and recognition, using the learned knowledge. Over years, we have seen that rigorous statistical and mathematical theories and algorithms have been developed for addressing these problems and that many have attempted to apply them to solve numerous difficult real-world problems. Today, we are realizing that alas! this paradigm of automation has serious limitations and many practical computing systems are incorporating clever but heuristic user interventions to achieve their end goals. What is really missing is a systematic theoretical and practical treatise in addressing these user factors within the existing computational theories for realizing reusable robust computing solutions. This talk overviews these ideas by illustrating some examples from my research field: visual information computing, especially medical image analysis. My previous and current studies of one-click data segmentation and registration, conducted at Siemens towards this goal, will be described and demonstrated briefly. I will also describe their foundation: how kernel-based robust statistical theories and algorithms can be applied to and extended for realizing these solutions.