Plans for the week of March 17-21

Dear all, welcome back to FYS5429/9429.

We hope obviously that the week has started the best possible way!  

This week the plan is to continue our discussions of Autoenconders (AEs) and linking it with the principal component analysis (PCA) method, which is a very useful and much used method for reducing the dimensionality of a given data set (finding the relevant features). 

We will show, for a linear transformation in an AE, that this corresponds to performing a PCA analysis.  The lecture will be devoted to a discussion of the PCA and its link with AEs. We will also show code examples in Pytorch and TensorFlow on how to use AEs.

The jupyter-notebook this week, with codes and more is at https://github.com/CompPhysics/AdvancedMachineLearning/blob/main/doc/pub/week9/ipynb/week9.ipynb

We recommend also revisiting the notebook from last week at 

https://github.com/CompPhysics/AdvancedMachineLearning/blob/main/doc/pub/week8/ipynb/week8.ipynb.

The notebooks contain also reading and coding suggestions, such as (you don't need to go through all obviously ?!). 

Reading recommendations

1. Goodfellow et al chapter 14.

2. Rashcka et al. Their chapter 17 contains a brief introduction only.

3. Deep Learning Tutorial on AEs from Stanford University at http://ufldl.stanford.edu/tutorial/unsupervised/Autoencoders/

4. Building AEs in Keras at https://blog.keras.io/building-autoencoders-in-keras.html

5. Introduction to AEs in TensorFlow at https://www.tensorflow.org/tutorials/generative/autoencoder

6. Grosse, University of Toronto, Lecture on AEs at http://www.cs.toronto.edu/~rgrosse/courses/csc321_2017/slides/lec20.pdf

7. Bank et al on AEs at https://arxiv.org/abs/2003.05991

8. Baldi and Hornik, Neural networks and principal component analysis: Learning from examples without local minima, Neural Networks 2, 53 (1989)

Best wishes to you all,

Edvin and Morten

p.s. In case you do only one project, please remember to upload by the deadline (March 21) a one-page summary of what you plan to do and what you have done till now.

For those of you doing two projects, let me know, if you an extension of the deadline.

p.s.2 If there are people doing two projects and in particular the CNN path as outlined in the project description, let us know if you are interested in more partners. Elias at elias.tidemandruud@gmail.com is still looking for additional partners.