Learning-based Dimensionality Reduction
Advanced techniques that will improve the applicability of adaptive sampling.
Dimension reduction is a crucial aspect of machine learning and predictive modeling. It has gained traction with the surge in data and the multifarious ways in which sensors are used to capture data and store it for subsequent analysis. Dimension reduction helps reduce data redundancy by eliminating unwanted dimensions that have no incremental value. There are several conventional methods to achieve dimensionality reduction, but they suffer from inherent limitations. To overcome such shortcomings, EPFL scholars are engaged in developing advanced techniques that will improve the applicability of adaptive sampling by absorbing the required information directly from data.
Jointly funded by ONR Global and Hasler Stiftung, the project is being conducted by PhD students Rabeeh Karimi and Arda Uran, and postdoc researcher Junhong Lin. They are investigating the broad areas where dimensionality reduction is relevant, with emphasis on key application areas such as medical imaging and array signal processing. These core areas are likely to see an increasing synergy between machine learning and human intuition. For instance, medical images are extremely complicated and high-dimensional, which creates major problems for analytical models. The researchers will propose novel techniques of dimensionality reduction to simplify medical imagery (such as neuroimaging) and yet preserve essential details.
They will also study the application of dimensionality reduction for computational tasks involving the use of sensor arrays. It is well established that computational complexity increases with an increase in the number of sensors in the array. To circumvent this problem, the researchers will look at new techniques to reduce the number of computations.
The current research veers away from other similar studies by adopting an interdisciplinary approach toward dimensionality reduction. It seeks to strengthen the relationship among allied disciplines, which include computer graphics, data mining, statistics, machine learning, signal processing, and optimization. This holistic approach is likely to lead to new findings related to sampling theory and methods.
The credentials of the researchers are in line with the project’s requirements. Junhong Lin has a keen interest in statistical learning theory and high-dimensional statistics, while Rabeeh Karimi and Arda Uran are pursuing their doctoral programs in Computer and Communication Sciences and Microsystems and Microelectronics, respectively.