One of the reasons I decided to give Python one last try is that Python is popular for machine learning, and machine learning is a topic I'm interested in.

Why?

Why am I interested in machine learning?

• It's an area of computer science that has advanced significantly since I studied CS in school
• It has many unique practical applications (speech recognition, language translation, machine vision, generative art)
• I recently acquired a GPU that (as far as I know) is capable of accelerating machine learning pipelines

First steps

As an introduction, I'm following MIT's Introduction to Machine Learning (2020) class. It's in Python and builds on NumPy. The first 4 weeks focus on linear classifiers for binary classification.

Although my math is rusty, my biggest struggle is actually with the NumPy API. Here is my list of grievances:

• NumPy implicitly "broadcasts" arrays into compatible shapes
  • I assume this is for convenience, but all it's done for me is silently hide bugs in my code
• NumPy sometimes silently removes dimensions
  • Again, this is probably for convenience, but all it's done is trip me up -- I really wish keepdims defaulted to True
• The API has a lot of cruft (e.g. a matrix class that is no longer recommended for matrix computation)
• The documentation is frustratingly vague, and sometimes circular (e.g. "* returns self * value")

Experimentation

Using a sample data set

To get a better handle on NumPy, I'd like to actually attempt to create a linear classifier from scratch. A quick search led me to a page with links to data sets for binary classification problems. I'm using UCI Machine Learning Repository's "banknote authentication" data set because the data format is simple (4 predictor variables and a 0 or 1 for the classification).

To my surprise, simple algorithms (e.g. selecting random parameters) were able to correctly classify over 95% of examples. For what it's worth, my code is posted here.

Solving a problem from scratch

I'd like to try solving a real world problem from scratch, but I don't really have a problem in mind that lends itself to binary classification. For what it's worth, here are some Kaggle data sets that might eventually inspire me:

• NBA players
• Hacker News data
• Chord classification
• Movie data
• Stack Overflow data