Massive data analytics in constrained settings
The ACES lab researchers work on creating disruptive automated methods, algorithms, and systems that serve the present demands of constrained embedded systems and enable novel types of data-intensive applications. While in recent years, machine learning (ML) and data-analytic experts have made impressive progress in adapting the algorithms to the geometric structure of the data, concerns over scalability and ease of use present roadblocks to the efficient adoption of these methods on constrained platforms. Our philosophy is that disruption in complex data-intensive embedded systems research requires foundational understanding of the form and behavior of the content and algorithms, as well as their interaction with the underlying hardware.
The ACES lab created the ﬁrst comprehensive data and platform-aware framework that provably optimizes the performance cost of learning analysis on dense data. The devised algorithms and hardware-customized engines reach the theoretical bounds on the system performance (in terms of runtime, energy, and/or memory usage) given a data geometry and platform architecture. The framework is applicable to a myriad of resource-intensive iterative ML applications and extended to support single-pass streaming applications. The usage scenarios include learning and analysis of dense visual content on small platforms, which is a key enabler for next-generation applications in important domains including healthcare, social media, robotics, and autonomous vehicles. The work so far has focused on the two widely-adopted ML algorithms: kernel based learning (e.g., regular and penalized regression, power iteration) and deep neural network based learning and is being extended to many more. The ACES lab provides customized solutions for many/multi-core CPUs and clusters, and heterogeneous platforms with FPGAs or GPUs. Please check our publications for details.