Prototype Learning for Explainable Image Analysis

Carlos Santiago

Instituto Superior Técnico, Lisboa

Resumen/Abstract

Prototype learning is an interpretable approach to image recognition where classification is based on how similar a test image (or parts of it) are to learned prototypes. These prototypes often represent characteristic examples (or parts of examples) from the training data for each class and offer a means to justify the decision made by the model. The similarity is typically measured in a latent space learned by a neural network, where samples from the same class are expected to be close to their corresponding prototypes. In this talk, I will present an overview of this line of work, providing a comparison with traditional deep learning-based approaches, advantages and shortcomings, and identifying different directions that have been explored.

Curriculum ponente

Carlos Santiago is an invited assistant professor at Instituto Superior Técnico (IST) and a researcher at the Institute for Systems and Robotics. He has an extensive background in computer vision, machine and deep learning, and his research and contributions are focused on image and video analysis, including object classification, detection, tracking, or counting, with applications in diverse areas such as healthcare, smart cities and agriculture. He has 13 top international journal publications, including IEEE-TPAMI, IEEE-TIP, PR, and over 30 international conference papers, of which two were best paper awards and two honorable mentions. Additionally, he has been involved in 10 national and 3 international projects, including collaborations with Carnegie Mellon University. His teaching activities include computer vision and big data processing courses at IST, and teaching post-graduate students on applications of artificial intelligence in Transportation.