Blog

Notes on machine learning, mathematics, and research.

Grokking Beyond the Euclidean Norm of Model Parameters

Why grokking is governed by the property favored late in training—and why that property need not be the Euclidean norm.
GrokkingRegularizationSparsityLow-Rank RecoveryImplicit Bias

What Is Grokking? A Formal Definition

A formal definition of grokking based on the delay between memorization and generalization.
GrokkingGeneralizationOverfittingOptimization

Smoothness, Descent, and Cocoercivity

Definitions, lemmas, and proofs connecting smoothness, descent guarantees, Hessian bounds, and the Baillon–Haddad theorem.
OptimizationSmoothnessConvex AnalysisCocoercivity

Epoch-wise Bias–Variance Decomposition

A bias–variance decomposition for tracking how bias, variance, and irreducible noise evolve throughout training.
deep learningstatistical learningbias-variance tradeoff

Visualization of the loss landscape and optimization path of a neural network

How low-dimensional subspaces can preserve an optimization trajectory and make a neural network’s loss landscape interpretable.
deep learningloss landscape

Word embeddings

The genesis of my word embeddings tutorial, and what led me to machine learning research.
NLPGloveWord2VecBag of wordsTF-IDF