[2D5-GS-2] Machine learning: applications (1)

[2D5-GS-2-03] Tsuchigumo: Training and Inference Deep Learning Framework with Posit

〇Masato Kiyama¹, Motoki Amagasaki¹ (1. Kumamoto University)

Keywords:Deep Learning, Deep Learning Framework

Quantization is effective to reduce the memory footprint, hardware resources, and power consumption of deep learning computations on edge devices such as AI chips.Posit number System is introduced for replacing IEEE standard 754 floating-point numbers (FP32).Posit is an effective numeric representation that can represent a value with less bit width than FP32 and can be used for quantization.We need a deep learning framework with posit to train and inference or to verify operations on edge devices.We developed Tsuchigumo, a DNN framework that uses posit in training and inference with the same behavior as hardware.In this paper, we describe an implementation of Tsuchigumo.Evaluations show that our framework works properly in training and inference using various posit formats.We evaluated overhead for emulation times and found that it took 3.5 seconds to train 100 data on MNIST.