The course is concerned with computational problems in massively parallel artificial neural network (ANN) architectures, which rely on distributed simple computational nodes and robust learning algorithms that iteratively adjust the connections between the nodes heavily using the available data samples. The learning rule and network architecture determine specific computational properties of the ANN. The course offers an opportunity to develop the conceptual and theoretical understanding of computational capabilities of ANNs starting from simpler systems and progressively studying more advanced architectures, and hence exploring the breadth of learning types – from strictly supervised to purely explorative unsupervised mode. The course content therefore includes among others multi-layer perceptrons (MLPs), self-organising maps (SOMs), Boltzmann machines, Hopfield networks and state-of-the-art deep neural networks (DNNs) along with the corresponding learning algorithms. An important objective of the course is for the students to gain practical experience of selecting, developing, applying and validating suitable networks and algorithms to effectively address a broad class of regression, classification, temporal prediction, data modelling, explorative data analytics or clustering problems. Finally, the course provides revealing insights into the principles of generalisation capabilities of ANNs, which underlie their predictive power.
Course structure
12 lectures, 4 labs, 1 project and exam
Course literature
- Stephen Marsland. Machine Learning, an Algorithmic Perspective, 2009,CSC-Press.
- Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning., 2016, MIT press.
Further recommended reading will be announced on the course's website.