Neural Networks

Faculteit Science and Engineering
Jaar 2020/21
Vakcode WBAI028-05
Vaknaam Neural Networks
Niveau(s) bachelor
Voertaal Engels
Periode semester II b

Uitgebreide vaknaam Neural Networks
Leerdoelen At the end of this course, the student:
- does not identify 'neural networks' with 'deep learning', but is instead aware of the interdisciplinary diversity of neural network modeling approaches in the wider neural, cognitive and information processing sciences
- can practically use feed-forward neural networks in machine learning applications
- understands elementary mathematical formalism to capture dynamical and stochastic phenomena in recurrent neural networks
- has an orientation in the wider research landscape of NNs which enables him/her to approach research fields adjacent to AI / machine learning, in particular computational neuroscience, theoretical physics / complex systems and neuromorphic microchip technologies.
Omschrijving 'Neural Networks' (NN) is an umbrella term for models of distributed, parallel information processing. In all neural network models, a (large) number of 'neurons' interact by spreading out activation patterns along 'synaptic' connections which can be shaped by learning mechanisms. This general scheme has been worked out into an enormous variety of concrete models in a diversity of scientific disciplines and application scenarios. This course gives an introduction to a choice of NN specimen which together illustrate the powers of NN models as an integrative link between AI, machine learning, computational neuroscience, theoretical physics and – since a few years increasing in importance – microchip technologies and material science. Concretely, the following topics will be treated:

- A rehearsal of machine learning basics: overfitting, regularization, cross-validation
- Perceptrons (classical, multilayer), and a glance at deep feedforward networks, such as CNNs and auto-encoders
- An analysis of computational challenges inherent to backpropagation
- A crash course on dynamical systems - to aid with the topic of Recurrent Neural Networks
- LSTM and other gated networks for "deep" learning of temporal data; backpropagation through time algorithm
- Hopfield networks – the classical NN model of neural memory and pattern completion
- Boltzmann machines and restricted Boltzmann machines
- Abstract 'NN-related' models of parallel information processing and self-organization: spin glass systems, Markov random fields, sampling algorithms
- A selection of (R)NN topics from computational neuroscience (to be decided), for instance cortical microcircuits, Bayesian information processing in spiking RNNs, neural motor control architectures, self-organizing maps, self-organized criticality or other (self-)regulatory mechanisms for neural dynamics, dendritic information processing
- Reservoir computing
- An introduction to 'neuromorphic computing', a field between NNs, microchip fabrication and material science, of particular interest in Groningen, as the core research theme of our CogniGron research center
Uren per week
Onderwijsvorm Hoorcollege (LC), Practisch werk (PRC)
(The practical work will be a practical miniproject, in parallel with the lecture, where a neural network presented in the course is implemented and applied to a task suitable for that kind of network. The deliverable is a final report. Weekly homeworks are for exercise only (non-graded).)
Toetsvorm Practisch werk (PR), Schriftelijk tentamen (WE), Verslag (R)
(Final Mark: The examination will count for 50% of the final mark and the final practical report also for 50%. The grade of the exam needs to be higher or equal to 5.0 in order to pass this course.)
Vaksoort bachelor
Coördinator prof. dr. H. Jaeger
Docent(en) prof. dr. H. Jaeger
Verplichte literatuur
Titel Auteur ISBN Prijs
A detailed and fully self-contained set of online lecture notes will be
made available.
prof. dr. H. Jaeger
Entreevoorwaarden Mandatory: Calculus (WBAI048-05), Linear Algebra and Multivariable Calculus (WBAI050-05).

If the mandatory requirements are not met, only the Board of Examiners of the AI BSc may grant an exemption. Exchange students are assumed to have gone through this through their Learning Agreement; pre-master's students through the Board of Admissions - other external students are judged case-by-case.
Opmerkingen This course was registered last year with course code KIB.NNKI03

Artificial Intelligence (BSc) is a Fixed Quota (Numerus Fixus) programme. As a consequence, their courses (course code WBAI) are closed for students that are not registered under the AI BSc programme, unless the course is part of the mandatory curriculum of their programme. If you wish to take this course in your minor – or as part of a so-called ‘unofficial’ pre-master’s – please use the official procedure through the Board of Examiners form.
Opgenomen in
Opleiding Jaar Periode Type
BSc Artificial Intelligence 2 semester II b verplicht
BSc Courses for Exchange Students: AI - Computing Science - Mathematics - semester II b