Titulo Estágio
Continual Learning for Evolving Image Recognition Tasks
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
CISUC
Enquadramento
Continual Learning (CL), also known as Lifelong Learning, is a field of machine learning focused on training models that can learn continuously from new data without forgetting what they learned before. In many real world applications, like species monitoring, new types of data appear over time. For example, new bird species that were not present in the original training set. Instead of training a new model from scratch every time new data becomes available, CL methods aim to update the model gradually, using only the new data.
A major challenge in this process is catastrophic forgetting, where a model forgets previously learned knowledge when it is trained on new information. This internship will focus on a specific type of continual learning scenario called class incremental learning, where new categories (or classes) are introduced over time, and the model must learn to recognize them without losing its ability to classify earlier ones. The goal is to explore and test different strategies for building a computer vision system that can learn in this way, using deep learning models like CNNs. A suitable dataset for this purpose is the “200 Bird Species with 11,788 Images” dataset available on Kaggle (https://www.kaggle.com/datasets/veeralakrishna/200-bird-species-with-11788-images).
Objetivo
The following objectives will guide the development of the audio-based classification system:
Study the state of the art in Continual Learning (CL) and machine learning
Study the available machine learning architectures, such as CNNs
Understand and test different CL techniques that help models learn new classes without forgetting old ones
Build a simple framework to run experiments using image classification models, such as CNNs
Start with standard datasets to compare methods
Apply the best methods to a realistic dataset, like bird species images, to test how well they work in more complex, real world scenarios
Plano de Trabalhos - Semestre 1
Conduct a literature review on class incremental continual learning in computer vision
Study CNN architectures commonly used in CL research
Set up benchmark experiments with datasets
Implement and test selected CL strategies
Write intermediate report
Plano de Trabalhos - Semestre 2
Adapt the CL framework to the bird image dataset
Evaluate different methods under incremental class arrival scenarios
Analyze performance and generalization under real world complexities
Propose possible improvements
Write final report
Condições
This work should take place in the context of a research project in CISUC. There is the possibility of a 6-month scholarship.
Orientador
Catarina Silva, Bernardete Ribeiro, Dinis Costa
catarina@dei.uc.pt 📩