Ch 3 — Machine Learning Paradigms — Under the Hood

Loss functions, gradient descent, K-Means internals, Q-learning math, and self-supervised pretraining
Index← High Level
Under the Hood
-
Click play or press Space to begin. Click any node for deep-dive details...
Step- / 10
ASupervised Learning InternalsLoss functions · gradient descent · evaluation
1
functions
Loss FunctionsMSE, cross-entropy
measuring error
2
trending_down
Gradient DescentOptimizing weights
step by step
assessment
EvaluationAccuracy, F1
train/test split
3
arrow_downward From labeled to unlabeled: unsupervised internals
BUnsupervised Learning InternalsK-Means math · PCA eigenvectors · anomaly detection
hub
K-MeansLloyd’s algorithm
iterative assignment
4
compress
PCAEigendecomposition
variance maximization
warning
AnomalyIsolation forest
outlier detection
5
arrow_downward From patterns to actions: reinforcement learning math
CReinforcement Learning InternalsMDPs · Bellman equation · Q-learning
schema
MDPStates, actions
transitions, rewards
6
calculate
Q-LearningBellman equation
value iteration
balance
ε-GreedyExplore vs
exploit balance
7
arrow_downward The bias-variance tradeoff and model selection
DModel Selection & ValidationBias-variance · cross-validation · regularization
tune
Bias-VarianceUnderfit vs
overfit tradeoff
8
grid_view
Cross-ValidationK-fold reliable
evaluation
shield
RegularizationL1, L2, dropout
prevent overfit
9
arrow_downward Modern paradigms: self-supervised & transfer learning
EModern ParadigmsSelf-supervised pretraining · transfer learning · RLHF
auto_awesome
Self-SupervisedNext-token, masked
contrastive learning
10
merge
Full PipelinePretrain + finetune
+ RLHF