Öffentliche Probevorlesung im Rahmen des geplanten Habilitationsverfahrens
Herr Dr. Julien Vitay
Fakultät für Informatik
Professur Künstliche Intelligenz
"Deep Learning Architekturen für Objekterkennung"
10:00 Uhr, Straße der Nationen 62, 1/336
Alle interessierten Personen sind herzlich eingeladen!
Classical computer vision techniques heavily rely on machine learning algorithms to solve complex tasks, including object recognition which associates a label to any possible variation (illumination, orientation, size, shape) of a single visual object such as a chair. These classical methods traditionally use extensive pre-processing steps to build a high-level representation of an image, which is then sent to a simple classifier.
In 2012, an end-to-end neural approach ("deep learning") was presented at the ImageNet object recognition challenge and outperformed the classical approaches by a factor two. Since then, object recognition almost uniquely rely on deep neural networks, attracting media coverage and industrial investments from major IT players. This lecture will present the reasons for the success of deep learning architectures in object recognition, building on the classical multi-layer perceptron to explain recent concepts such as convolutional layers, batch normalization and dropout. It will also present unsupervised approaches to object recognition, including stacked restricted Boltzmann machines, and will propose solutions to the race towards always deeper neural architectures.