bayesian-dropouts
Project for Bayesian statistics implementing the methods from Y. Gal and Z. Ghahramani, “Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning,” Apr. 2016
Data and code summary
We expiremented on one regression tasks and two classification tasks:
-
regression : PM2.5 concentration in China, dataset available here. The regression task is conducted both with the gaussian process and with the dropout NN models in regression_GP.ipynb and regression_NN.ipynb.
-
classification 1: MNIST (we used keras example datasets: from keras.datasets import mnist) implemented in classification_MNIST.ipynb.
-
classification 2: NotMnist, a similar dataset as mnist, slightly more complex on 10 characaters (available here). We downloaded the notMNIST_large.tar.gz (originally 600 000 samples) and cut it to 10000 images per classes using the file not_mnist_creation.py. The classification is implemented in classification_notMnist.ipynb.
Utils files implement the useful functions that we use for the regression and classification tasks.
Instructions
- (1) explain the theoretical, computational and/or empirical methods,
- (2) emphasize the main points of the paper,
- (3) apply it to real data (that you will find).
Bonus points will be considered if you are creative and add something insightful that is not in the original paper: this can be a theoretical point, an illustrative experiment, etc.
Format
You can use either Python or R for the programming part. Please have each group send
- one report as a pdf (≤ 15 pages, with reasonable fonts and margins),
- one zipped folder containing your code and a detailed readme file with instructions to (compile/install and) run the code.
to all three teachers 1 no later than January 30th. There will be no deadline extension.