Evaluating classification models for resource-constrained hardware

• Ano de defesa: 2020
• Autor(a) principal: Lucas Tsutsui da Silva
• Orientador(a): Gustavo Enrique de Almeida Prado Alves Batista
• Banca de defesa: Anna Helena Reali Costa, Ronaldo Cristiano Prati, Roseli Aparecida Francelin Romero
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade de São Paulo
• Programa de Pós-Graduação: Ciências da Computação e Matemática Computacional
• País: BR
• Link de acesso: https://doi.org/10.11606/D.55.2020.tde-11112020-180216

Machine Learning (ML) is becoming a ubiquitous technology employed in many real-world applications in diverse areas such as agriculture, human health, entomology, and engineering. In some applications, sensors measure the environment while supervised ML algorithms are responsible for interpreting these data to make an automatic decision. Generally, these devices face three main restrictions: power consumption, cost, and lack of infrastructure. Most of these challenges can be better addressed by embedding ML classifiers in the hardware that senses the environment. Thus, we need highly-efficient classifiers suitable to execute in unresourceful hardware. However, this scenario conflicts with the state-of-practice of ML, in which classifiers are frequently implemented in high-level interpreted languages (e.g., Java or Python), make unrestricted use of floating-point operations and assume plenty of resources such as memory, processing and energy. In this work, we present a software tool named Embedded Machine Learning (EmbML) that implements a pipeline to develop classifiers for low-power microcontrollers. This pipeline starts with learning a classifier in a desktop or server computer using popular software packages or libraries as WEKA or scikit-learn. EmbML converts the classifier into a carefully crafted C++ code with support for resource-constrained hardware, such as the avoidance of unnecessary use of main memory and implementation of fixed-point operations for non-integer numbers. Our experimental evaluation on benchmark datasets and a variety of microcontrollers shows that EmbML classifiers achieve competitive results in terms of accuracy, classification time, and memory cost. Compared to classifiers from some existing related tools, ours achieved the best time and memory performances in at least 70% of the cases. Lastly, we conduct experiments in a real-world application to describe the complete pipeline for using EmbML and assessing its classifiers with an intelligent trap to classify and capture flying insects.