Studentische Arbeiten

Titel: Deep learning for channel coding
Typ:
  • Masterarbeit
Betreuer:
Status: laufend

Hintergrund

During the last decade, machine learning and deep convolutional networks have achieved superhuman performance in visual domains, such as image classification and description, face recognition and playing of computer games. Recently, a deep learning-based computer program has beaten the European Go Champion, one of the artificial intelligence challenges considered to be at least a decade away. Similar breakthroughs happen in other domains such as health care and real-time speech translation; businesses are taking a new look at so-called cognitive technologies due to dramatic performance improvements in recent years.

One of the most complex tasks in wireless communications is channel decoding. For example, in LTE systems, the Turbo-decoder alone requires around 70% of the entire computing resources of a base station (eNB). The goal of this master thesis is to analyze the suitability of deep learning for the implementation of a channel decoder.  The expected benefits of this approach are a better performance and/or lower complexity than traditional algorithms.

Problemstellung

Deep learning-based channel decoding is doomed by the curse of dimensionality: for a short code of length N = 100 and rate r = 0.5, 250 different codewords exist, which are far too many to fully train any neural network (NN) in practice. The only way that a NN can be trained for practical blocklengths is, if it learns some form of decoding algorithm, which can infer the full codebook from training on a small fraction of codewords. In our recent work (see [1]), one of our key finding is that structured codes are indeed easier to learn than random codes, i.e., less training epochs are required. Additionally, our results indicate that NNs may generalize or “interpolate” to the full codebook after having seen only a subset of examples, whenever the code has structure.

The major objective of this thesis is a further investigate of possible network architectures such that they generalize better.

[1] https://arxiv.org/pdf/1701.07738v1.pdf

Aufgabe

•    Get familiar with existing Keras framework
•    Structured analysis of required training dataset
•    Investigation of different network structures with respect to their generalization capability
•    Thesis can be written in German/English


•     Methods
•    Keras (with Tensorflow/Theano backend)
•    Python programming (as Jupyter notebook)
•    GPU-aided learning

Anforderung

•    Good knowledge of channel coding
•    Basic knowledge of machine learning techniques (deep neural networks, stochastic gradient descent)
•    Experience with Python, C/C++
•    Experience with machine learning frameworks is a plus (e.g.,Tensorflow, Caffe, Theano/Lasagne, Torch)

Kenntnisgewinn

•    Introduction into deep-learning and neural networks
•    Work with state-of-the-art machine learning tools (Keras, Theano, Tensorflow,...)
•    Gain basic knowledge about polar codes