just ppt

Created using ChatSlide

The research focuses on On-Device Human Activity Recognition (HAR) Training, exploring its purpose and the variety of techniques utilised. A comprehensive framework for classifier architecture is provided, detailing a 15-layer model and its equations. Quantisation techniques are examined, including STE-only, NIPQ, and fixed-point methodologies. Solutions to identified software issues such as memory bugs and NIPQ bit-packing are discussed. Measurement techniques include allocator spy and size-...