Nik Tavakolian, Busra Tas Kiper: Modern Approaches for Representation Learning with Image data

Abstract

Representation learning (RL) is a machine learning paradigm characterized by autonomous extraction of features (representations) from raw data that are useful for some downstream tasks. For image data, RL is an important step in successfully performing tasks such as image classification and enhancement (e.g. generating super-resolution images from old pictures). In this talk, we cover the intuitions (instead of technical details) of some modern approaches of RL.

In the first part, we present RL in the context of image classification when the correct classes for most images are unknown. In classification one wants to classify a set of images into a number of discrete classes. Therefore, one constructs a representation where images in the same (different) classes have similar (distinct) representations. In the second part, we focus on the concepts behind image super-resolutionization such as pixel interpolation and sparse coding. Finally, we briefly cover Convolutional Neural Networks (CNNs) and their applications in image processing. Specifically, we explore the ideas of Super-Resolution CNNs that aim to increase the resolution and sharpness of images.