Deep Learning for Media Technology

Aim

The course teaches fundamental techniques in deep learning, theoretical as well as practical, and gives an overview of modern techniques and applications related to deep learning for images, graphics and sound.

After finishing the course, the student should be able to:

• describe fundamental techniques for how to construct and optimize artificial neural networks, and show understanding of how theoretical concepts relate to practical situations,
• demonstrate knowledge on modern methods for how neural networks can be designed, optimized and used in different contexts within computer vision, image processing, computer graphics, natural language processing and visualization,
• use existing deep learning tools for solving classification and and regression problems for digital media,
• use techniques for testing and improving the performance of neural networks, including concepts such as detection of overfitting and techniques for increasing generalization capability,
• formulate and solve simpler problems from start, including data collection, selection of techniques, and analysis of the results,
• reason around the consequences of an increasing use of deep learning in the society, both in terms of opportunities as well as problematic questions.

Content

Overview of machine learning. Artificial neural networks (ANN) and the term deep learning. Learning paradigms (supervised/unsupervised/semi-supervised/self-supervised/reinforcement learning). Optimization of ANNs (back-propagation, stochastic gradient descent, momentum, batch normalization). Regularization (augmentation, drop-out, early-stopping). Data (images/video/sound/3D, representation, training/testing, bias, adversarial examples). Architectures (convolutional networks, auto-encoders, recurrent networks, recursive networks). Generative deep learning. Applications (computer vision, image processing, computer graphics, natural language processing, visualization) and implications (societal impacts, ethics, bias).