Para receber por e-mail novas publicações de Seminário de Computação Gráfica, Clique Aqui!

Seminários do IMPA

Computação Gráfica

Título
On the Visual Integration of Training and Unseen Data in Classification
Expositor
Bruno Schneider

University of Konstanz
Data
Quarta-feira, 29 de agosto de 2018, 13:30
Local
Auditorio 3
Resumo

There is a growing interest in the field of machine learning, which enables a computer to learn from data instead of explicitly programming it to execute a particular task. Classification is one of the problems in machine learning. Examples go from automatically recognizing images after training a model with known instances, supporting the diagnosis of diseases using medical records, or categorizing our e-mail messages as spam or not spam.

One of the problems in Classification is that we never know if the performance obtained during the construction of a model will be the same with new and unseen data. To better understand the reasons for poor model generalization, I propose the visual integration of training and unseen data. In this work, we want to explore, understand and explain how the lack of similar learning examples affects the classification outputs with unseen data at the time of training.

I will also show my previous work on the visual integration of data and models in Ensemble Learning. The goal, in this case, is to give direct access to models and data spaces in classification, thus enabling the user to explore the relationships between these spaces and seek for classification patterns that are not visible through aggregated model performance metrics.