Analytical guarantees for reduced precision fixed-point margin hyperplane classifiers

Fixed-point machine learning

Sakr, Charbel ; Shanbhag ; Naresh R.

关键词: Fixed-point machine learning;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/99324/SAKR-THESIS-2017.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Margin hyperplane classifiers such as support vector machines are strong predictive models having gained considerable success in various classification tasks. Their conceptual simplicity makes them suitable candidates for the design of embedded machine learning systems. Their accuracy and resource utilization can effectively be traded off each other through precision. We analytically capture this trade-off by means of bounds on the precision requirements of general margin hyperplane classifiers. In addition, we propose a principled precision reduction scheme based on the trade-off between input and weight precisions. Our analysis is supported by simulation results illustrating the gains of our approach in terms of reducing resource utilization. For instance, we show that a linear margin classifier with precision assignment dictated by our approach and applied to the `two vs. four' task of the MNIST dataset is ~2x more accurate than a standard 8 bit low-precision implementation in spite of using ~2x10^4 fewer 1 bit full adders and ~2x10^3 fewer bits for data and weight representation.