A notebook on statistical learning
Ideas, in the
margins.
Papers on statistical learning, read through their questions, methods, and limits. Follow the ideas; keep the original sources close.
The notebook
08 NOTESReusing a Transformer block
Shared parameters, repeated computation, and task-dependent benefits.
Depth as a fixed point
An equilibrium representation separates model definition from its numerical solver.
When shared parameters improve estimation
A reading of the public abstract: parameters, iteration, and accuracy.
What a coverage guarantee means
Calibration turns a score into a prediction set with a stated guarantee.
What generated scores tell us about uncertainty
Reading GSI through its coverage guarantee and generator requirements.
How to read a statistical guarantee
A practical way to connect a theorem’s assumptions, probability statement, and conclusion to the question you actually want to answer.
What makes a comparison fair?
A good benchmark starts with a clear question, an appropriate evaluation design, and an explicit account of the resources each method uses.
From a result to reproducible evidence
Build a clear path from the research claim to its inputs, analysis, and outputs, so another person can inspect and repeat the work.
A way of reading
Less noise.
More understanding.
Start with the question
Name the problem before reaching for a method.
Stay close to the evidence
Keep assumptions and conclusions in the same frame.
Leave the limits visible
A useful result also tells us where it stops.