MACHINE LEARNING
Applied Machine Learning Notes
Mainly these notes are for me to refer back to, but I’ll make this public just in case it helps other Medium subscribers
I LOVE free courses. It’s crazy to me that everybody everywhere isn’t obsessed with learning future tech. For free. On demand. By the world’s top professors.
Why do anything else?
I’m currently enjoying a course by Columbia University Professor and Scikit-Learn core developer Andreas Mueller even more than one that I recently started by the legendary Stanford Professor and Coursera co-founder Andrew Ng. This is because it’s more focused on application rather than the underlying theory and mathematics.
Both instructors are fantastic, and I do plan on going back to Professor Ng’s CS229 lectures for more depth, and as a co-requisite for CS230. But since I’m simultaneously trying to learn Python, this Machine Learning course so far seems ideal for my needs.
Before attempting any of these courses, I’ve already taken the beginner course CS50 by Harvard’s David Malan as well as a 4-hour python bootcamp by freeCodeCamp.com. I’ve also been practicing routinely on, well, Practice Python. I highly recommend both.
I will keep updating this article with additional notes as I go, so it may get quite long.
As a side note, I’m not writing because I think I’m already an expert. Precisely the opposite; I’m writing because I think the best way to learn something is to teach it. It forces me to learn each concept for myself in a way that can be explained simply.
Again, I’m just a beginner, so I deeply appreciate any corrections or suggestions more experienced computer scientists might offer.
Applied Machine Learning 2020 by Andreas Mueller
Lecture 1
Video Course Schedule Python Data Science Handbook by Jake VanderPlas (free textbook)
Main Types of Machine Learning: 1) Supervised (spam detection, medical diagnosis, ad click prediction, etc…) 2) Unsupervised (pre-process for supervised learning, exploratory analysis) 3) Reinforcement Learning (Alpha Go, robotics. Agent vs Environment)
p(x,y), where x is data and y is a desired result f(x) = y, that’s the prediction goal when unknown x values introduced
Classification and Regression Classification is for discrete data, like finding if a tumor is present or not Regression is for continuous data, like patient recovery time
Relationship to Statistics Statistics finds the model first and has inference emphasis Machine Learning has data first and has prediction emphasis
Grey area overlap. Each one is an appropriate tool for different goals. Statistical Models: The Two Cultures by Leo Breiman (free pdf download)
Cost of Complex systems Don’t automatically rush to the most complex approach. Use a simple and cheap type where possible. Get baseline data before trying to improve something with machine learning.
Good and Bad Substitutes Usually the goal is to optimize for something too complex like profit, and so use something like “time on platform viewing ads” as an easy substitute for this which we can optimize for. However, it can produce unethical outcomes like radicalizing desensitized users with progressively more extreme content.
Explainable results make algorithms more accepted and less angering to users as long as reasons for things like recommendations or bans are plausible. It doesn’t actually have to be true. But lie to users at your own peril, because it’s hard to restore broken trust.
Ethical Considerations Everyone knows humans are prone to bias, and some people hoped that algorithms would provide an objective, unbiased and prejudice-free decision-making tool. Unfortunately it isn’t that simple. Machine learning predictions tend to amplify existing biases through a feedback loop rather than correct for it. Because of American history and demographics, in the United States many seemingly unrelated variables like income or zip code can actually correlate to and become unintended proxies for race. We have to be very careful and rigorous to avoid codifying further injustice into predictive models.
Free vs Expensive Data Predicting observable events tends to be free, as data for things like stock markets, clicks, and house numbers is readily available. But automating complex processes tends to be expensive.
Lecture 2
(coming soon… in the meantime I’m practicing the basics of Python with these resources)
To make sure you don’t miss content like this, simply become a member and subscribe!
