Predictive prognostic models based on transfer learning

• Grant number: 24/22546-0
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: April 01, 2025
• End date: March 31, 2028
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Lilian Berton
• Grantee: Manuel Alberto Bezerra Brandao Corrales
• Host Institution: Instituto de Ciência e Tecnologia (ICT). Universidade Federal de São Paulo (UNIFESP). Campus São José dos Campos. São José dos Campos , SP, Brazil
• Company: Secretaria de Desenvolvimento Econômico (São Paulo - Estado). Instituto de Pesquisas Tecnológicas S/A (IPT)
• Associated research grant: 20/09850-0 - Applied Artificial Intelligence Research Center: accelerating the evolution of industries toward standard 5.0, AP.PCPE

Abstract

In traditional supervised machine learning, the goal is to develop a model that becomes successful and efficient in a task based on example data. Typically, specific models are built that excel only in solving a single specific problem. This traditional development approach incurs a high time cost to create, train, and fine-tune models from scratch for each new problem, even if it is a problem that has already been addressed in the industry. Transfer learning aims to address this issue by leveraging existing knowledge and data from one domain to apply it to a new one. Several researchers emphasize that transfer learning will be essential for commercial and business success. Andrew Ng, a renowned professor and data scientist who has been associated with Google Brain, Baidu, Stanford, and Coursera, stated in 2016: *"After supervised learning, transfer learning will be the next driver of commercial success."* Transfer learning has proven particularly effective in image data, where it is common to utilize a deep learning model trained on large image datasets, such as ImageNet. These pre-trained models can be directly integrated into other new models that take images as input. In textual data, pre-trained models also exist for learning word representations. These can be embedded into deep learning language models at both the input and output stages. The objective is to develop predictive prognostic models for a specific industrial sector that can be partially reused as a starting point to predict failures in another sector of the company using transfer learning. To achieve this, features between the source and target tasks will be mapped, and techniques will be developed to enhance learning in target tasks by leveraging knowledge from the source task. This approach aims to optimize learning time and improve the accuracy of the target task. (AU)