Learning Track

Stats to Deep Learning

You have a statistics background. This path connects classical stats to modern deep learning. It translates what you already know into ML language, then shows you what's genuinely new and what's familiar in disguise.

1/5

Phase 1: Stats You Already Know (in ML Language)

You know these concepts from stats. Here they are in the ML context with the notation that papers use.

MLESame as stats, but now for neural networks Bias-VarianceSame idea, different failure modes in high dimensions Fisher InformationThe Cramér-Rao connection to natural gradient Hypothesis Testingp-values for model comparison BootstrapResampling for uncertainty, same as in stats

2/5

Phase 2: What's Different in ML

This is where classical stats intuition starts to break. High dimensions change everything.

Concentration InequalitiesThe tools that replace CLT in finite samples ERMThe ML analog of MLE PAC LearningWhen is a problem learnable?Cross-ValidationModel selection without test-set abuse RegularizationRidge and lasso, but with a learning theory lens

3/5

Phase 3: Neural Networks for Statisticians

How neural networks relate to function estimation, loss minimization, and nonparametric methods.

Neural NetworksComposition of nonlinear functions Universal ApproximationWhy neural nets can fit anything (in theory)Activation FunctionsReLU, sigmoid, GELU DropoutStochastic regularization, Bayesian connection OptimizationHow gradient descent actually works

4/5

Phase 4: Deep Learning Specifics

The machinery of modern deep learning. Transformers, attention, and the training pipeline.

TransformersThe dominant architecture AttentionContent-based weighting, not convolution Scaling LawsPower laws in model performance Batch NormWhy it helps training, not why you think Weight InitializationXavier, He, and why random matters

5/5

Phase 5: Where Stats and DL Reconnect

Modern theory that bridges classical statistics and deep learning.

Double DescentBias-variance is incomplete CalibrationPredicted probabilities vs actual accuracy Bayesian Neural NetworksPosterior over weights Gaussian ProcessesThe kernel connection Implicit BiasWhy overparameterized models don't overfit

Strong in stats, new to ML?

Phase 1 maps what you already know. Phase 2 is where the new stuff starts. If Phase 1 is too easy, jump straight to Phase 2.