As the field of computer vision has rapidly been overtaken by deep learning, computer vision researchers have stepped to the forefront, both in pioneering and applying innovative deep-learning methods. Yet there are some problems for which deep learning is either impractical or suboptimal. Researchers equipped with both deep learning skills and an understanding of key classical methods/techniques will find themselves best prepared for the increasingly visual future of data science.
In this five-session course, I will present a carefully selected combination of classical and deep learning methods for a variety of visual data problems. We will learn together about the following topics:
- Introduction to Visual Data - Visual data comes from diverse sources, has myriad representations and can be analyzed in many different ways.
- Advanced Clustering and its Applications - Clustering is an important method for dealing with big data, but there are many nuances in how this is best done.
- Metric Learning and Alternate Representations - Particularly with high-dimensional data, distance metrics, including those learned by supervision, can help us to distinguish the impediments from the important parts of our data.
- Data Visualization - In this era of deep-learning, one of the most important things that a data scientist can do is to spend time looking deeply at their data.
- The Categorization Spectrum - In this final session, we will discuss the spectrum of recognition, beginning with basic-level objects like cars, trees and people and focusing in on the often long-tailed/data-limited problems of fine-grained recognition and biometrics.