Talk Summary: Machine Learning (ML) and in particular Deep Learning (DL) require huge numbers of MIPS to implement the training and classification processes required of modern Edge AI applications. All of this comes at a cost of MIPS , power consumption and silicon cost. The frequency domain and, in particular, the Fourier Transform can be used to pre-process the data in many of todays Edge-AI applications to reduce the number of MIPS required for both training and inference. This paper discusses how the frequency domain algorithms can be used and how they are not quite as frightening as they at first appear.
Bio: John Edwards is a DSP and Embedded Systems Consultant. He has worked as a DSP Engineer since the early 1980s, including wireless and digital communications, control, automotive, IoT and AI, working for companies such as Loughborough Sound Images, Motorola, Picochip and XMOS Semiconductors. Since 1993 John has been a visiting lecturer at the University of Oxford, presenting the 4 day Digital Signal Processing course as part of the annual Summer Engineering Program for Industry. He is a member of the IET, IEEE and regular contributor at international DSP conferences.