Assignment 4

Total marks: 5
Use TFP with Jax substrate for distributions
Use Jax for autograd functionality and vector programming
The assignment has to be done in groups of two
The assignment should be a single Jupyter notebook.
It is important that you are able to get insights from your results and not just present answers. As an example – if the question asks you to try a different likelihood, then you should be able to observe how this likelihood compares to the default likelihood.

For all of the following questions, please consider binary classification as the model.

Use a subset of 5,000 images (500 images per class) from MNIST dataset and train a simple (deterministic)CNN model in Flax. Report the predicted accuracy on a separately kept 1,000 images test set (100 images per class). Also, report the predicted probabilities for wrongly classified examples. What can you conclude? Now, use the trained model on a set of 1,000 images on Fashion MNIST dataset. What can you conclude? [1.5]
Modify the above CNN to include MC dropout as way to get uncertainty from the final dense layer. Rerun above experiments and report conclusions. [2]
Use a GP library of your choice and kernel of your choice to fit a GP to monthly CO2 data from Mauna Lua. Compare the prediction with a neural network that uses MC Dropout to obtain uncertainty. [1.5]