Prof. Bart ter Haar Romeny, Eindhoven University of Technology
The Masterclass on Deep Learning will not only discuss the layers of the convolutional networks and the application of modern Deep Learning tools, but will go further. We will discuss the intrinsic mechanisms of the black box: the essential notion of representation learning, and the neuro-mathematics of self-organization and contextual processing. We will take a deep tour into modern vision and brain research, both in the retina as visual cortex, and realize the strong similarities between the visual pathway and Deep Learning, and how much be learned from one by studying the other. We demonstrate surgical software tools to dissect the network layers and look inside what they compute, and explain modern visualization tools, like t-SNE.
All essential mathematics is explained in an intuitive way, with many visual and hands-on examples.
The masterclass can be followed by all interested in deep learning, with an emphasis on trying to understand what happens inside. All code of all examples is given to the participants.
09:10 – 10:00 Overview of Deep Learning today. Machine learning, application areas, Python/TensorFlow/Keras/Mathematica, learning resources. Backpropagation, MNIST
10:00 – 11:00 Convolutional Neural Networks I
Convolution, template matching, regression, detailed layer descriptions
11:00 – 11:15 Break
11:15 – 12:15 Convolutional Neural Networks II
Data preparations, data augmentations, residual neural networks, long short-term memory networks.
Applications: Retinal image analysis, face recognition, inference of trained neural nets
12:15 – 13:30 Lunch
13:30 – 14:30 Understanding what the first layers do Geometrical model, first principles, representation learning, unsupervised learning, contextual processing
14:30 – 15:30 Modern human vision and brain research Brain imaging, the multi-modal retina, cortical columns, pinwheels, visual pathways, feedback
15:30 – 16:00 Break
16:00 – 16:45 What happens inside the black box? Surgery on network layers, layer visualizations, t-SNE, face representations
16:45 – 17:30 New ideas: learning network efficiency from human vision modeling Video processing, color analysis, foveation
Conclusion, overview/summary of the day