Module 6 Review
You have practiced communicating technical work, reading research and testing an idea from a paper. This page sums up the Self-study of this module and checks the main ideas.
Summary
- Communicating Technical Work: principles for presentations and for writing, turned into your own checklists. A document is useful when it is valuable to its reader.
- Understanding ML Research: types of research output, authorship and peer review, the scientific method, searching the literature, and why published results depend on the setting in which they were obtained.
- Research to Practice: choose a principle from a paper, apply it to a project, and evaluate the change under your usual validation.
Quiz
These questions are easy if you have worked through the activities. Choose an answer to see whether it is right and why.
According to the lecture on writing, what matters most in a technical document?
- That it is complete and covers every detail. A complete document can still be useless to its reader.
- That it is valuable to the reader. Right. Clear, organized and persuasive writing only helps if the content is valuable to the reader.
- That it uses the technical terms of the field. Terms help when the reader knows them, but they are not what makes a document useful.
You look for papers about a general method, such as hyperparameter optimization. How do you start?
- Broad searches first, then more specific ones. Right. The broad searches give you the words for the specific ones.
- By searching for the exact title of the best paper on the topic. You usually do not know that title yet; finding it is the goal of the search.
- By reading only the most cited paper. Citations show attention, not that a paper fits your problem.
A highly cited paper reports that a technique improves results. What do you do before you rely on it?
- Nothing: many citations show that the technique works. Results depend on the data, the models and the evaluation in the paper, as the SMOTE example shows.
- Use it only if the paper is recent, because newer methods replace older ones. Newer is not always better; what matters is whether it works in your setting.
- Check its setting, and test it on your own problem. Right. Collect your own evidence under your own validation.
You test an idea from a paper in your project. How do you evaluate it?
- Compare with and without the change on the same validation folds. Right. Then a difference comes from the change, not from other folds.
- Look at the test score after the change, because it is the most honest score. Choosing on the test score makes it too optimistic.
- Check whether the authors used the same dataset as you. That is useful background, but it does not evaluate the change in your project.