Machine Learning¶

Machine Learning is the field of enabling computer systems to learn and improve from experience without being explicitly programmed. It's the key to building flexible, adaptive intelligent agents.

Learning Problem Definition¶

A learning agent improves its performance based on: - Experience: Examples from training data - Performance metric: How well it achieves the task - Task: What the agent is trying to learn

General framework: - Collect training examples: (input, output) pairs - Choose a hypothesis space (possible functions) - Use a learning-algorithm to find best hypothesis - Test on new data (generalization)

Learning Paradigms¶

Confusion Matrix for Classification Confusion matrix showing true positives, false positives, true negatives, and false negatives. Used to evaluate classification model performance.

Supervised Learning¶

Learn function from labeled examples (input → output) - Regression: Continuous output (price, temperature) - Classification: Discrete output (category, yes/no) - Requires labeled training data - Examples: neural-networks, decision trees, SVM

Unsupervised Learning¶

Find patterns in unlabeled data - Clustering: Group similar examples - Dimensionality reduction: Find lower-dimensional structure - No target output specified - Examples: k-means, principal component analysis

Reinforcement Learning¶

Learn via interaction and reward/punishment - Agent takes actions, receives rewards/penalties - Goal: maximize long-term cumulative reward - No labeled target outputs - Examples: Q-learning, policy gradient methods

Semi-supervised Learning¶

Combination of labeled and unlabeled data - Use small labeled set + large unlabeled set - Often more practical than pure supervised

Key Concepts¶

Bias-Variance Tradeoff¶

Bias: Error from oversimplified model (underfitting)
Variance: Error from model sensitivity to training data (overfitting)
Tradeoff: Increase complexity → decrease bias, increase variance
Regularization: Add penalty term to prevent overfitting

Generalization¶

Learning system must generalize to new, unseen data
Training error ≠ test error
Cross-validation: Estimate test error using multiple train/test splits

Hypothesis Space¶

Set of possible functions the learner can express
Small space: Can't learn complex concepts
Large space: Risk overfitting, harder to search
Inductive bias: Preference for simpler hypotheses (Occam's-Razor)

Sample Complexity¶

How many examples needed to learn well? - Depends on: task complexity, hypothesis space size, desired accuracy - VC-Dimension: Measure of hypothesis space complexity

Learning Algorithms¶

Passive Learning¶

Learn model from data, then use it (no exploration) - Examples: Decision trees, neural-networks, Bayesian learning

Active Learning¶

Learner chooses which examples to learn from - Often more efficient (fewer examples needed) - Examples: Uncertainty sampling, query by committee

Online Learning¶

Learn continuously as new data arrives - Single pass through data - Update incrementally - Examples: perceptron, gradient-descent variants

Batch Learning¶

Collect all data first, then learn - Can optimize globally - Requires storing all data

Performance Evaluation¶

Metrics¶

Accuracy: Fraction of correct predictions (classification)
Precision/Recall: For imbalanced classes
Mean squared error: Average squared difference (regression)
F1 score: Harmonic mean of precision and recall

Validation Strategies¶

Train/test split: Use 70/30 or 80/20 split
K-fold cross-validation: Split into k parts, test on each
Leave-one-out: Extreme case: test each single example

Neural-Networks — Nonlinear model learned via backpropagation
Decision-Trees — Interpretable tree-based models
Gradient-Descent — General optimization method
Reinforcement-Learning — Learning via rewards
Bayesian-Learning — Probabilistic approach to learning

References¶

Russell & Norvig (2010): Chapter 18 - Learning from Examples