What ML researchers and users get wrong: optimistic assumptions

ML is often done poorly, both by “ML experts” (by which I mean people who understand the algorithms but not the data) and “ML users” (by which I mean people who understand their data, but not the algorithms). I think the cause is often over-optimism, although about different things:

• When designing algorithms, ML experts make optimistic assumptions about the data (eg it is IID, lies on a low-dimensional manifold, has sub-Gaussian noise)
• When selecting algorithms for their data, ML users make optimistic assumptions about behavior of the algorithms (eg it is robust to outliers, will discover hidden “truths” in data, will always be better than predicting the mean)

In reality, these assumptions are usually unwarranted. It is helpful to keep in mind:

• ML experts: there is a whole universe of crazy data out there. Some data is so noisy that it’s basically just noise. Some data is aggregated from 100 heterogeneous sources. Data will have random labelling artifacts (eg, some survey respondents will pick A for every answer just to complete the survey quickly). Every time you add an assumption to your analysis, you reduce the class of problems your method can be applied to. If you don’t document these assumptions clearly, people will misuse your algorithms.
• ML users: AI is not magic and is often more brittle than you think. Algorithms can “overfit” to your data. LLMs make things up. Test set loss only estimates generalization if drawn from the same distribution you will see at deployment (among other conditions). I like this image from scikit-learn which shows how common clustering algorithms can give weird results, even on simple toy datasets.