Improving prototypical networks performance on few-shot image classification with genetic programming

• Grant number: 22/06772-4
• Support Opportunities: Scholarships abroad - Research Internship - Scientific Initiation
• Start date: July 01, 2022
• End date: September 28, 2022
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: João Paulo Papa
• Grantee: Giovani Candido
• Supervisor: Javier Del Ser Lorente
• Host Institution: Faculdade de Ciências (FC). Universidade Estadual Paulista (UNESP). Campus de Bauru. Bauru , SP, Brazil
• Institution abroad: Fundación Tecnalia Research & Innovation, Spain
• Associated to the scholarship: 20/16092-5 - Fine-tuning prototypical neural networks using metaheurístics, BP.IC

Abstract

Over the years, supervised image classification has made some noble advances. The classification models, in turn, developed a stronger need for larger datasets. However, in many situations, images are scarce. In this context, few-shot learning stands out as it allows models to learn from a few labeled samples. In particular, a promising model known as Prototypical Networks was introduced. This model can learn a metric space where similar images gather, while dissimilar images distance themselves from one another. In this space, a prototype is computed with the mean of all feature vectors belonging to its class. Then, any test image can be classified by finding the prototype nearest to its feature vector. As the model assumes each image has the same level of importance in the computation of the class prototype, leading to bad prototypes in some cases, the Weighted Prototypical Networks were proposed. This other model employs a three-layer neural network to find weights for each training sample and take these into account when obtaining the prototype. However, since it is desired to improve the performance with the weights, it is naturally an optimization task where the goal is to minimize the classification loss. Under this assumption, the present project hypothesizes that employing meta-heuristic techniques would be a better course of action, compared to the three-layer neural network. To test this hypothesis, the project intends to use Genetic Programming, once it is well-known and related to the main research project of the student. Additionally, it is worth mentioning that the approach would be validated under the supervision of Prof. Javier Del Ser Lorente, the main researcher operating in the areas of data science and optimization at Tecnalia, which is one of the biggest research centers in Spain. Moreover, to validate the proposed approach, it would make use of not only general-purpose data sets but of endoscopic images as well, so that we could distinguish Barrett's esophagus from adenocarcinoma. (AU)