Active learning model with neural networks for applications with costly labeling

• Grant number: 21/11058-6
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: April 01, 2022
• End date: August 31, 2022
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Odemir Martinez Bruno
• Grantee: Gustavo Vieira Jodar
• Host Institution: Instituto de Física de São Carlos (IFSC). Universidade de São Paulo (USP). São Carlos , SP, Brazil
• Associated research grant: 18/22214-6 - Towards a convergence of technologies: from sensing and biosensing to information visualization and machine learning for data analysis in clinical diagnosis, AP.TEM

Abstract

Neural networks and machine learning have emerged in recent decades, both in academia and industry. Commercial solutions can be found today ranging from medicine with X-ray image recognition, to predictions of consumer decision-making for the industry. However, although the purposes are numerous, the development of accurate models is hampered by the constant need to label training data, which can be an undesirable bottleneck for many applications with costly labelings, such as protein classification, and image classification in computer vision and problems involving large data sets. Thus, the area of active learning in machine learning introduced techniques to reduce the need for labeling in training data, the most used methods are part of the Pool-Based Selective Sampling group, in which it is assumed that there is a data set not labeled, and it is necessary to select specific samples from this set, by some measure of informativeness, and label them for training. However, there are flaws in the main approaches of this group and, even the most complex ones, require labeling in the chosen dataset. The model proposed by this scientific initiation project aims, after initial training, essential in any scenario, to remove the need for labeling in the data, using dimensionality reduction techniques, classifying algorithms, and various neural networks. Recently, the candidate carried out preliminary tests with the proposed model in a simplified form, which proved to be promising for their results and for favoring various fields of research if the study of the subject is further studied.(AU)