Prep for David Rolnick, Part 1

Part A (Reading)

First, please spend approximately 30 minutes reading this position paper by Dr. David Rolnick and a large team of coauthors (Rolnick et al. 2024).

Optional extra: Dr. Rolnick also suggested this “huge” (his word) paper on climate change and ML as of possible interest for many of you, but we agreed that it was too big to expect you to read in preparation for our time with him.

Part B (Comprehension Check)

Please write 1-2 sentences in response to each of the following prompts. Some of the questions below are explicitly designed to require some light reflection or investigation beyond the text of the assigned reading.

  1. What is application-driven machine learning research, and how (according to the authors) does it differ from methods-driven machine learning research?
  2. Land-use classification, identification of animals from photographs, and labeling of medical images of tumors as benign or malignant are all, from a methods perspective, examples of image classification. In what ways do the appropriate evaluation metrics differ across these three applications?
  3. In today’s discussions of generative AI models, we often talk about foundation models. What is a foundation model? How does the idea of a foundation model compare or contrast to the paradigm of application-driven machine learning research?
  4. We often hear about the brilliance of “advanced” researchers in AI and ML who are exceptionally strong in mathematics, algorithms, and programming. We less-frequently hear about “advanced” researchers who are exceptionally strong in collaboration with domain experts, design of new evaluation techniques, or development of new methods for data collection. What are some of the obstacles mentioned in the reading that lead to the first set of skills often being labeled as more “prestigious” than the second?
  5. What are three reasons that the pace of publication in application-driven machine learning research is often slower than the pace of publication in methods-driven machine learning research?

Part C (Fix My Class)

In responding to this question, one piece of possibly-helpful context is that I am not a machine learning researcher (application-driven or otherwise).

Reflect back on the first half of this course, CSCI 0451: Machine Learning. Would you describe the development of the course so far as more methods-driven or more application-driven? Can you give examples of each approach in what we’ve covered so far?

Based on the reading, please suggest one or more ways in which an aspect of this course could be redesigned to be more application-driven.

Part D (Question for Dr. Rolnick)

As you know, we will have the chance to talk with Dr. Rolnick in class on Monday, 3/30. Please write a question that you would like to ask him in class!

This can be a question about this reading (or anything else by him that you’ve found), Dr. Rolnick and his research program, his career path, the state of machine learning and AI, the intersection of CS and climate change, or anything else that you are curious to get an expert opinion on.

References

Rolnick, David, Alan Aspuru-Guzik, Sara Beery, Bistra Dilkina, Priya L. Donti, Marzyeh Ghassemi, Hannah Kerner, et al. 2024. “Position: Application-Driven Innovation in Machine Learning.” In Forty-First International Conference on Machine Learning. https://openreview.net/forum?id=xEB2oF3vvb.