The lecture Machine Learning for the Natural Sciences promises to focus on applications of machine learning to natural sciences, especially physics and chemistry. However, most of the actual content is repeating machine learning basics, that is already in foundational lectures on machine learning. In the remaining time, a few interesting are presented, but sadly just very shallowly.
There is also programming homework that counts for 1/3 of the final grade. This is nice, and I think more courses should do that. Such homework has the potential for much more learning than studying for the final exam. However, the programming exercises here are filling in blanks in jupyter notebooks of mixed quality and difficulty.
The final exam fulfills almost all Exam Anti-Patterns. Especially not publishing old exams. So here are the questions of the exam I can still remember.
- Given a picture of 2D space with data points. What split achieves perfect purity?
- Calculate Bayes Rule: Population of 99% farmers and 1% librarians. 5% of farmers enjoy Sci-Fi and 90% of librarians enjoy Sci-Fi. Given someone enjoys Sci-Fi what are approximate probabilities that they are a farmer or a librarian.
- Which plot shows the ReLU function?
- What metric is good for testing for a rare but very deadly disease → Sensitivity
- What is the maximum entropy and when is it reached
- Advantages of transfer learning
- Advantages of CNNs
- How many learnable weights does a polling layer have
- By how much is an image decreased after going through a convolutional layer with some parameters and through pooling with some other parameters afterwards.
- Design an ML pipeline for a scenario: Predicting band gaps from SMILES with some labeled data and a simulation tool
- What is the Markov property (but there are two answers correct, which is unexpected)
- Draw example SMILES code and explain how it works in general
- How do molecule fingerprints work? Are they useful as output of a generative model?
- Some multiple choice question about the target networks in deep q learning
- What are the two limitations of deep q learning and what can we do about that
- Do we need a readout function for a GNN when used for learning potentials of single atoms in a molecule
- How to represent molecules in a GNN (some weird question with hand wavy terminology)
- What does the radius in a fingerprint correspond to in a GNN
- What is attention and why is it useful for translation and chemical reaction prediction
- Explain the Bayesian learning algorithm and two applications in ML and the natural sciences for it
- Explain a possible idea for parallel Bayesian learning
- Query by committee: What is it and how is it used in enhancing the test set