Scheda di revisione: Fundamentals of Machine Learning

Machine Learning Revision Sheet

1. 📌 Essentials

• Machine learning creates models that predict outcomes based on data.
• Models are functions y = f(x), trained on labeled datasets.
• Features (x) are input variables; labels (y) are targets.
• Types include regression, classification, clustering, and deep learning.
• Regression predicts continuous values; classification predicts categories.
• Deep learning uses neural networks with layered architectures- Model evaluation uses metrics like MAE, MSE, accuracy, F1-score.
• Training involves data to derive the function; inference predicts new data.
• Supervised learning uses labeled data; unsupervised learning finds patterns without labels.
• Common algorithms: Linear regression, logistic regression, K-Means, neural networks.

2. 🧩 Key Structures & Components

• Training Data — past observations with features and labels.
• Features (x) — input attributes, e.g., temperature, measurements.
• Labels (y) — target output, e.g., sales, species, risk.
• Model — the learned function y = f(x).
• Algorithm — method to fit data, e.g., gradient descent.
• Neural Network — layered structure mimicking neurons for deep learning.
• Loss Function — measures prediction error during training.
• Evaluation Metrics — assess model performance (e.g., accuracy, R2).

3. 🔬 Functions, Mechanisms & Relationships

• Data → features (x) and labels (y) → model training.
• Training adjusts model parameters to minimize loss.
• Model predicts ŷ (predicted label) from new features x.
• Regression models output continuous y; classification models output probabilities.
• Deep learning uses multiple layers to extract complex patterns.
• Evaluation metrics guide model tuning and selection.
• Hierarchical flow: Data collection → preprocessing → training → validation → deployment.
• Supervised models require labeled data; unsupervised models find inherent data structure.
• Neural networks perform forward pass, compute loss, then backpropagate errors to update weights.

4. 📊 Comparative Table

Item	Key Features	Notes / Differences
Regression	Predicts numeric y; continuous output	Example: sales based on temperature
Binary Classification	Predicts 0/1; probability-based; evaluated with confusion matrix	Example: disease risk prediction
Multiclass Classification	Predicts among multiple classes; softmax or OvR	Example: penguin species
Clustering	Unsupervised grouping; no labels; based on similarity	Example: customer segmentation
Deep Learning	Neural networks with multiple layers; backpropagation	Example: image recognition, NLP tasks

5. 🗂️ Hierarchical Diagram

Machine Learning
 ├─ Data
 │    ├─ Features (x)
 │    └─ Labels (y)
 ├─ Model
 │    └─ Function y = f(x)
 ├─ Training
 │    ├─ Fit algorithm to data
 │    └─ Derive model parameters
 ├─ Inference
 │    └─ Predict ŷ from new x
 ├─ Types
 │    ├─ Regression (numeric y)
 │    ├─ Classification (categorical y)
 │    │   ├─ Binary (2 classes)
 │    │   └─ Multiclass (>2 classes)
 │    └─ Clustering (unsupervised)
 └─ Deep Learning
      └─ Neural networks, layered architecture, backpropagation

6. ⚠️ High-Yield Pitfalls & Confusions

• Confusing regression (numeric y) with classification (categorical y).
• Overfitting: models perform well on training but poorly on new data.
• Misinterpreting probability outputs as final predictions in classification.
• Using accuracy alone in imbalanced datasets; consider precision, recall.
• Assuming deep learning is always necessary; simpler models often suffice.
• Mixing up supervised and unsupervised methods.
• Ignoring feature scaling, which affects algorithms like K-Means.
• Overlooking the importance of validation and test datasets.
• Confusing model evaluation metrics; e.g., R2 vs. accuracy.

7. ✅ Final Exam Checklist

• Understand the difference between supervised and unsupervised learning.
• Know the main types of models: regression, classification, clustering.
• Be able to describe features, labels, and how models are trained.
• Recognize common algorithms: linear regression, logistic regression, K-Means, neural networks.
• Know how to evaluate models: MAE, MSE, R2, accuracy, precision, recall, F1-score.
• Understand the training process: forward pass, loss calculation, backpropagation.
• Be familiar with deep learning architecture and its applications.
• Know examples of real-world applications for each model type.
• Understand the importance of data quality and feature engineering.
• Be aware of common pitfalls like overfitting and bias.
• Recognize the role of evaluation metrics in model tuning.
• Know platforms like Azure ML for deployment and scaling.

End of Revision Sheet