Some notes and questions based on Breiman’s article before
These are some notes for a fairly impromptu journal club for the Independent Summer Statistics Community in the Department of Statistical Sciences at the University of Toronto. The idea is to chat about the original article before going to a talk called ‘Learning from Data: The Two Cultures’. (Info/registration). The speaker is Adji Bousso Dieng (Founder, The Africa I Know; Google AI; Princeton University (Incoming)) and moderator is David Blei (Columbia University; ACM Prize in Computing Recipient).
Breiman’s article, free and in full: http://www2.math.uu.se/~thulin/mm/breiman.pdf.
Have the paper open (even if you haven’t had a chance to read it all) Introduce yourself: Name, year & program of study, why you were interested in this talk.
This talk is part of a series of ‘TechTalks’ from the Association for Computing Machinery (ACM). You can view past ACM TechTalks on their learning portal. They also have a podcast series, ByteCast.
You’ll see the idea of ‘Two Cultures’ in STEM related topics pop up now and then and (I think) they call back to a lecture given by C. P. Snow in 1959 (Wikipedia). His talk (and later book) was about how the sciences and the humanities has become two separate ‘cultures’ and that this split was a problem when it comes to trying to solve our big problems.
Right at the start of the introduction, Breiman says there are two goals for traditional statistical analyses: prediction and information. In your studies so far, have you done both, one more than the other, only one?
Are you more naturally a data modelling culture member or an algorithmic modelling culture member?
Writing: Did you find this article easy or hard to read? What did you like/helped? What was difficult?
In the description of the Ozone project (Section 3.1), how did the way the training and test sets were set up differ to what you might have seen usually done? Why do you think this was? Could the normal way also have worked? (Note: if you’ve never encountered training/test sets, let’s talk about that!)
In section 3.3, Breiman discusses his perceptions about statistical analysis after year of consulting, upon going back into academia. Did any of these surprise you? Do you disagree with any? Which of these perceptions do you already hold?
In section 5.2, when residual analysis is helpful/not for understanding model fit. Was there anything new you learned from this section?
Section 5.3, discusses ‘the multiplicity of data models’. What does this mean to you? Can you connect it to the idea ‘all models are wrong, but some are useful’?
Breiman set up the below as his goals for the paper. Do you think they were achieved?
Focusing just on the data modelling culture has:
Led to irrelevant theory and questionable scientific conclusions;
Kept statisticians from using more suitable algorithmic models;
Prevented statisticians from working on exciting new problems;
What is one thing in the article you’d like to learn more about and how? (Self-learning, in your courses, at future optional events?)
In the summary for the upcoming talk, it is stated that “there is a growing community of researchers working on methodologies embracing both cultures”, but goes on to identify two NEW cultures and how they need to work together. Any predictions about this topic? Examples from your own experience?
“If our goal as a field is to use data to solve problems, then we need to move away from exclusive dependence on data models and adopt a more diverse set of tools”
“One is left with an unsettled sense about the arbitrariness of residual analysis”
“Occam’s Razor, long admired, is usually interpreted to mean that simpler is better. Unfortunately, in prediction, accuracy and simplicity (interpretability) are in conflict”
“the emphasis needs to be on the problem and on the data.”
Display photo by freestocks on Unsplash
For attribution, please cite this work as
Bolton (2021, July 13). Liza Bolton: Learning from Data: The Two Cultures (journal club). Retrieved from blog.lizabolton.com/posts/2021-07-12_two-cultures/
BibTeX citation
@misc{bolton2021learning,
author = {Bolton, Liza},
title = {Liza Bolton: Learning from Data: The Two Cultures (journal club)},
url = {blog.lizabolton.com/posts/2021-07-12_two-cultures/},
year = {2021}
}