Weeekly FYS-STK3155/4155 update

Dear All, here comes (delayed) the weekly update with plans for the week.

As you may have seen, the project deadline has been extended by a week, with deadline October 8. This means that today's lab session will focus only on project 1. Project 2 will be available next Tuesday and will focus on combining classification and regression using what you have developed in project 1 as well as logistic regression and neural networks.

This leads us to this week's lecture plan:

1) Thursday we will derive the back propagation algorithm and discuss how to build our code for a multilayer perceptron model for deep learning. We will also discuss coding of the algorithm. This is a central algorithm in neural networks. You can find all the material in the slides, see for example https://compphysics.github.io/MachineLearning/doc/pub/NeuralNet/html/NeuralNet-bs.html and read from slide 21 and on. You can also run interactively the jupyter notebook if you prefer, go then to https://github.com/CompPhysics/MachineLearning/blob/master/doc/pub/NeuralNet/ipynb/NeuralNet.ipynb or just do a git pull and have everything on your own laptop/pc.

Neural networks are also discussed in chapter 11 of Hastie et al.

2)

We will also discuss how to use tensorflow and on Friday we will have a more hands-on session during the lecture. Tensorflow is one of the more popular Python packages for deep learning. Bring your laptops with you. Project 2 will also use tensorflow. You find more info in the above lecture slides, with examples etc and the text of Geron, see chapters 9 and on of  Hands?On Machine Learning with Scikit?Learn and TensorFlow, O'Reilly, or just grab it from the github address at https://github.com/CompPhysics/MachineLearning/blob/master/doc/Textbooks/TensorflowML.pdf

Finally, I must confess that I got really frustrated trying to understand the back propagation algorithm described in the textbook of Hastie et al, see chapter 11.4.

If you are able to derive the back propagation algorithm based on equations 11.8 to 11.15 in an understandable way, I'll offer the first of you a book reward at Akademika bookstore worth 250 NOK.

Best wishes for the week,

Morten et al