English

If sufficient demand

The study circle introduces doctoral students to recent advances in deep learning. Focus is on using deep learning in practice.

The content is based on the doctoral students' choice and active participations. Possible subtopics include Autoencoders, Convolutionan Networks, Structured Probabilistic Models, Restricted Boltzmann Machines, Recurrent and Recursive Nets, Deep Reinforcement Learning, Tensorflow, Deep Learning using GPUs, Stacked Denoising Autoencoders, DL for Natural Language Processing, Deconvolutionan networks, Optimization of Deep Networks.

understand how deep networks can be constructed and trained.
know typical components of a deep network such as RBMs, autoencoders, RNNs.
understand how typical network and training parameters influence the performance of the networks and their training speed.

be able to construct and train a deep neural network using an existing software platform, such as Tensorflow, Theano or similar.
be able to understand how to map a practical problem to a deep network architecture.
read and understand research articles in the field of deep learning.
be able to choose efficient network training methods.

Lectures

Laboratory exercises

Written assignments

Seminars given by participants

For a passing grade the doctoral student should give a lecture on a deep learning topic and construct a suitable hand-in assignment for the other course participants. The student should also solve at least half of the exercises presented in the course and upload solutions and code to a common code repository.

Failed, pass

The participants are assumed to have taken a course in basic machine learning or have corresponding knowledge.

Goodfellow, B. & Courville: Deep Learning (manuscript).

Book manuscript: https://github.com/HFTrader/DeepLearningBook/blob/master/DeepLearningBook.pdf

FRT170F

 -06-01

Professor Thomas Johansson