Skip to main content
— Presentation at IEEE ICASSP 2016 —
April 13, 2016
The 41st IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP 2016) was held by the IEEE Signal Processing Society in Shanghai, China during 20-25 March 2016. ICASSP is the world's largest international conference on signal processing. A lot of research results on wide areas are presented in ICASSP every year. For example, research topics are not only audio and speech signal processing such as speech enhancement and speech recognition but also signal processing theory, machine learning for signal processing, signal processing for wireless communication, bio-signal processing, etc. ICASSP 2016 received a total of 2682 paper submissions, and 1265 papers were accepted for presentation (Acceptance rate: 47%). More than 2300 participants discussed about the cutting edge of signal processing.
Fig. 1 Proposed approach
In ICASSP 2016, Hitachi, Ltd. Research and Development Group made a presentation titled "Adaptive Boolean Compressive Sensing by Using Multi-Armed Bandit". This presentation is about a method for solving "group testing". Group testing is a problem that appears in many areas such as genetic screening, anomaly detection in network, substance monitoring in wide-area, etc. So, a method for solving group testing can be also used to many applications. The goal of group testing is to find a small number of anomalies from among a lot of samples. If each sample is tested individually, a large number of tests will be necessary. Therefore, in group testing, to reduce the number of tests, multiple samples are mixed into one mixed-sample, and the mixed sample is tested instead of testing each sample individually. By repeating these steps changing the combination of mixing, a sequence of test results is obtained. Then, based on the sequence of test results, the anomalies can be found through only a small number of tests. However, there was a problem that, if the number of the samples selected for mixing is not controlled accordingly to the number of the anomalies, the required number of tests will increase. It is difficult to control the number of the samples selected for mixing because the number of anomalies is cannot known in advance. The proposed method is a combination of "compressive sensing", which has been studied in signal processing area, and "multi-armed bandit", which is an approach of reinforcement learning has been studied in artificial intelligence area. The proposed method can control the number of the samples selected for mixing adaptively and find the anomalies through only a small number of tests.
(By KAWAGUCHI Yohei)