Cynthia Rudin @ ICBINB Seminar Series - 2022

Title : Cynthia Rudin @ ICBINB Seminar Series Author(s): I Can’t Believe It’s Not Better! Link(s) : https://www.youtube.com/watch?v=XZUZ87wfq70

This is a talk/discussion on the importance of applications, title of the talk is "Applications Really Matter (And Publishing Them Is Essential For AI & Data Science)".

Cynthia starts by mentioning her experience in predicting power failures in NYC (, ), which humbled her and made her realize her perspective on ML at the time were wrong. In a real world project like this, the data was messy, and there were really high stakes. There is also a belief, (or rather, a religion), that having elegant math drives performance. Also, there is an informal concensus that interpretability has no value, and at the time atleast people did not believe that more powerful ML didn't do anything on such problems.

Applied science is science.

Talk outline:

• Benefits to science.
• Benefits to real world.
• Broadening our community.
• Freeing our top scientists.

Applications should be driving ML methods development. Some general papers on application based ML: (, a, cite:, a).

Some places that reject application papers include NeurIPS, AAAI, IEEE Transactions of Big Data, JAIR, Management Science, Nature Machine Intelligence. Some places where Cynthia was able to publish papers include Journal of Quantitative Criminology, Decision Support Systems, Physiological Management, TACL, Harvard Data Science Review (all of which are journals, which are slow).

TLDR, we cannot see how data science/ML is changing the world by looking at NeurIPS, ICML etc.

Cynthia claims that ML (currently) does not accomodate sufficiently diverse view of what constitutes top science, and applied work is somewhat considered to be second tier work.

Overall, this leads to researchers having to choose between working on problems that are either important to society and those that are beneficial for their own career based on how the publishing system currently works.

A solution to this is to open an applied track in all conferences/journals. The KDD applied track is not enough since it focuses more on deployment rather than science.