Year: 2020
Type: dissertation
Abstract: In recent years, Machine Learning (ML) went through a renascence due to improvements in computing systems and computer memories.The internet also played an important role, by providing access to and aggregating large amounts of data.As this technology evolves, optimizations to its processes are receiving more attention.Traditionally, machine learning models are intense in both memory and computations during training and inference.An optimization technique used in ML is focused on the inference phase.Models are typically trained in 32-bits, but instead of performing inference in 32-bits (operations and storage), it can be quantized to a format that uses fewer bits -this is called Post-training quantization.Usually, the fewer bits being stored and moved around in a computing system, the less energy is consumed, thus faster computations are performed, resulting in a more efficient system, given equivalent tasks.The goal of this study is to compare two 8-bit Post-training Quantization techniques by using two different basic models and exploit both their potentials and caveats.Both models are trained to classify handwritten numbers, the first one is focused on Fully Connected layers while the second focuses on Convolutional Layers.One of the techniques examined adopts a novel numeric representation system and this work also explores a model to understand how the system accumulates error.In short, it is an attempt at understanding which method provides a more efficient and practical solution.In both case studies, TensorFlow Lite is concluded to be the best Post-training Quantization solution.
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W3039946573
