Title: Deep Learning – beyond training with labels
Speaker: Vlado Menkovski (TU/e WSK&I)
Time: May 14, 2020, 10:00–11:00
Location: Online (MS Teams)
Abstract | Machine learning based solutions are being developed for a growing number of tasks in a broad range of fields. Deep Learning techniques have extended the scope of these solutions to high dimensional input spaces of images, natural text, and speech. It is still most common that these solutions take a supervised learning formulation, where for either observed or computed input data an expert provides a target value. In this talk I discuss the limitations of these approaches, particularly regarding understanding generalization capabilities of the model, interpreting the models’ behavior as well as utilizing efficiently the available expert knowledge.
I reflect on this through the experience of developing a model for classification of tokamak plasma confinement states. Going forward I evaluate alternative formulations for machine learning, focusing specifically on developments in deep metric learning and learning disentangled representations as a way to address the shortcomings of supervised learning.