Visual Analytics for The Design of Convolutional Networks

• Grant number: 25/21479-0
• Support Opportunities: Scholarships abroad - Research Internship - Doctorate
• Start date: March 01, 2026
• End date: February 28, 2027
• Field of knowledge: Physical Sciences and Mathematics - Computer Science
• Principal Investigator: Alexandre Xavier Falcão
• Grantee: Matheus Abrantes Cerqueira
• Supervisor: Alexandru-Cristian Telea
• Host Institution: Instituto de Computação (IC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Institution abroad: Utrecht University,
• Associated to the scholarship: 23/09210-0 - A Human-in-the-loop Approach to Build Convolutional Neural Networks, BP.DR

Abstract

Convolutional Neural Networks (CNNs) have highly improved image-pattern recognition tasks primarily due to their feature learning function (encoder), given a large data volume and available hardware scenario. However, the absence of designing methodologies under expert supervision often results in models that are more complex than necessary and whose internal functionality is unknown, reducing security and confidence in such models.Alternatively, Feature Learning from Image Markers (FLIM) incorporates an expert in the learning process by learning convolutional filters directly from marked patches, resulting in competitive and compact networks that require only a few marked images (1-10). However, the performance of this method is highly related to user actions/judgment, without an off-the-shelf solution or guidance. Therefore, we propose to develop a methodology to create FLIM networks more objectively, with assistance for users' actions. For this, in our main project (#2023/09210-0), we have studied solutions for interactively selecting images according to well-defined criteria and also investigated how the markers' positions impact the model's performance. The goal of this BEPE is to enhance the methodology by recasting the process of designing FLIM networks as an interactive and iterative Visual Analytics-enabled loop, creating tools to visualize and help the user during the marking process, and illustrating the feature learning process. This BEPE project will also lead to a joint PhD degree for the candidate at Utrecht University under the supervision of Prof. Alexandru Telea (the co-advisor). (AU)