Issue #19: How to read ML papers
(Without a PhD in mathematics )
ML papers terrify you.
You open a PDF. You see a wall of Greek symbols, dense equations, and references to theorems you’ve never heard of.
You close it within 30 seconds and you tell yourself:
“I’m not ready for this. Maybe once I learn more Math.”
That day never comes.
Here’s the truth: you don’t need a PhD to read ML papers.
I would recommend my 5-step process that takes you from “I can’t read this” to reading 2-3 papers per week.
Step 1: Stop reading top to bottom
This is the biggest mistake. Papers are not books and reading them linearly is the slowest way to understand them.
Instead, read in this order:
Title + abstract (30 seconds, decide if it’s worth your time)
Figures and tables (2 minutes, this is where the actual results are)
Last 2 paragraphs of the introduction (the contribution summary)
Method section (only the parts relevant to what you need)
Skip proofs entirely
80% of a paper’s value is in the abstract, figures, and method section. The rest is context for reviewers, not for you.
Step 2: Realize the math is the same every time
Here’s what most people don’t expect.
After 10-15 papers, the same operations appear over and over:
→ Matrix multiplication and transposition
→ Dot products and norms
→ Derivatives and gradients
→ Softmax and argmax
→ Summation and expectation
→ Argmin over a loss function
That’s it. That’s 80% of all Math you’ll encounter in ML papers.
You don’t need to know all of mathematics. You need to know these 10-15 operations deeply.
Step 3: Build a notation cheat sheet
Every time you see a symbol you don’t recognize, write it down.
Here’s what a typical cheat sheet looks like after ~50 papers:
→ θ = model parameters
→ L = loss function
→ ∇ = gradient
→ ∈ = “belongs to”
→ ∀ = “for all”
→ ||·|| = norm (length of a vector)
→ ~ = “sampled from”
→ E[·] = expected value
After 20 papers, you stop seeing new symbols. The notation becomes a language you can read fluently.
Step 4: Implement the key equations
This is the step that separates people who read papers from people who understand them.
After reading, ask yourself: “What are the 3-5 equations that define this method?”
Then try implementing only those using just NumPy.
Step 5: Start with the right papers
Don’t start with a 40-page theory paper. Start with these:
“Attention Is All You Need”
“Word2Vec: Efficient Estimation of Word Representations in Vector Space”
“ImageNet Classification with Deep Convolutional Neural Networks”
1 paper per week. In 3 months, reading papers becomes routine.
The uncomfortable truth:
ML papers aren’t written to be easy to read. They’re written to pass peer review.
But the actual ideas inside them? Usually simple,elegant,and implementable in 20-50 lines of code.
Although it seemed so, barrier was never your math background.
It was not having a system.
You don’t need a PhD to read ML papers.
You need a system: skip to the abstract and figures first, learn the same 10–15 math operations that repeat everywhere, build a notation cheat sheet, implement the key equations in NumPy, and start with accessible papers like “Attention Is All You Need.” 1 paper a week. In 3 months, it’s routine.
Thanks for reading this week’s issue.
Until next week’s issue, keep learning, keep building, and keep thinking like a mathematician.
-Terezija
P.S. If this issue helped you feel less intimidated by research papers, forward it to a friend who’s been avoiding them.
That’s how Math Mindset grows, 1 reader at a time.