Compressed sensing techniques for sensor data using unsupervised learning

Song will discuss compressed sensing, which is a powerful sparse data compression and recovery technique that is still growing fast since it was proposed in 2004. The data with certain sparseness are first compressed by using a dimension reduction method. The compressed data can be stored or sent to other devices. When there is a need to analyze the data, it can be reconstructed from its compressed version by least squares optimization with L1 regularization. This technique is extremely useful in sensor applications where data can be analyzed offline or on the cloud to alleviate power consumption, storage, transmission and computation limitations of the sensor. Examples include daily activity recognition, sports assessment, and patient monitoring mobile devices where large amount of data can be generated and battery power saving is desired. In addition, compressed sensing is also applied as a matrix completion tool to study recommendation systems.

In this talk, the basic theory of compressed sensing will be explained with simple numerical examples. Several case studies in wearable sensors, medical imaging and pattern recognition will be presented as well.

More details on compressed sensing can be found from a Stanford graduate course (An introduction to compressed sensing (http://www-stat.stanford.edu/~candes/stats330/index.shtml2)) website on the topic.

Speaker bio: Song Cui is a post-doctoral fellow of Molecular Imaging Instrumentation Lab at Stanford. His research interests include several machine learning topics such as nonlinear parameter estimation, advanced convex cost function design, time series analysis, clustering algorithm, and their applications to sensors and instrumentations. More details about the speaker can be found here (http://www.linkedin.com/profile/view?id=71703589).