Deep Learning: A Primer on the Algorithms Dominating Computer Vision

John McKay

Penn State University

In 2012 the computer vision community was rocked by a breakthrough paper from a group at the University of Toronto. Krizhevsky et al used a deep convolutional neural network (CNN) to blow away benchmarks in object recognition, almost halving the previous top error rates. From that point on, it has been almost impossible to avoid deep learning in computer vision circles. How did researchers take this once unappreciated model and make it into the dominant force in machine learning? We will cover deep learning and focus on modern strategies. In particular, we will dive into the concept of transfer learning where existing models are, with little or no modification, adapted to various domains. In settings where deep learning has been infeasible given limited data, transfer learning has emerged as a viable strategy and we will show how even CNNs trained on images vastly different than the transfer set can still yield compelling if not state-of-the-art results.

Back to ACMS schedule