enlargeme

Code I used for my Master Thesis: Super-resolution of magnetic resonance imaging data using machine learning tools.

Motivation

Master Thesis as mentioned, but mostly self-edu. Machine Learning became my main field of study in 2015 and since then I used popular frameworks like caffe, matconvnet or torch to build neural networks (mainly for image processing). This time I wanted to implement this structures myself to get more familliar with low level details. Also these are my first steps in unsupervised learning!

Solution

First version was to create two structures that can extract features from respectively low and high resolution images, and then create non linnear mapping between feature maps. After presenting a LR image to the first structure features would be extracted, then mapped to HR feature maps. Then SR image would be generated.

For now I am working on training RBM that will have real-valued visible layer (as its input is image) and convolutional connection between layers. When this part is done - after providing CCRBM (Convolutional Continuous RBM) an MRI image, and sampling v->h->v layer output will be at least as good as input image - non linnear mapping will be implemented. The plan is to use a neural network, but I'm not sure about the details. Should it be convolutional as well, or simple MLP will be enough.

TO DO:

• [x] Implement CCRBM model with activations
• [x] Implement contrastive divergence algorithm
• [x] Add batch version of CD
• [x] Implement Persistant CD algorithm for more efficient learning
• [x] Implement learning with scalar B and C values
• [x] Enable saving models to files and then loading it back to python
• [x] Implement methods for testing current solutions, visualise filters, monitor mean square error
• [x] Compare scalar B and C version with master where there are matrices
• [x] Try to combine both versions by first using scalar values and then matrices B and C (for more efficient learning)
• [ ] Implement momentum for contrastive divergence for faster filters training
• [x] Implement extra improvements mentioned in [2]
• [ ] Keep README.md up to date
• [ ] Make enalargeme a package that one can install with pip
• [ ] add some CI like Jenkins or Travis
• [ ] suggestions?

Recently I started zenBoard for this project and that is where futher imporovement can be found

Usefull information

[1] Fischer, Igel: An introduction to Restricted Boltzmann Machines (2012)

[2] Hinton: A Practical Guide to Training Restricted Boltzmann Machines (2010)

[3] Norouzi, Ranjbar, Mori: Stacks of Convolutional Restrited Boltzmann Machines for Shift-Invarian Feature Learning

[4] Lee, Largman, Pham, Ng: Unsupervised feature learning for audio calssification using convolutional deep belief networks

[5] Lee, Grosse, Ranganath, Ng: Convolutional Deep Belief Networks for Scallable Unsupervised LEarning of Hierarchical Representations

[6]Bengio, Delalleau: Justifying and Generalizing Contrastive Divergence (2007)

[7] Chen, Murray: Continuous restricted boltzmann machine with an implementable training algorithm (2003)