Tunable Floating-Point for Embedded Machine Learning Algorithms Implementation

The development of embedded and real-time systems for Machine Learning data processing is challenging (e.g. IoT). Low latency, low power consumption and reduced hardware complexity should be the characteristics of such systems. Considering prosthetic applications, which are error-tolerant, a technique that tunes the precision of operands and operations has been chosen for a Machine Learning algorithm used for tactile data processing. This paper presents the implementation of a Tunable Floating-Point (TFP) representation into a Singular-Value Decomposition (SVD) algorithm based on the One-Sided Jacobi method. The TFP representations demonstrate high performance and efficiency improvements of the SVD algorithm. © 2018 IEEE.