ML & DL

This module is organized primarily by learning signal. This avoids duplicating the same algorithm under multiple parallel directory trees. Other taxonomic perspectives are kept on this page to help readers understand cross-cutting relationships between models.

Recommended Resources

• ML & DL Resources

ML Algorithm Taxonomy

By Learning Signal

Learning signal describes what kind of information a model learns from.

• Supervised Learning

  • Learns a mapping from input to output using labeled data.
  • Data form: $\mathcal{D}=\{(x_i,y_i){\}}_{i=1}^n$
  • Learning target: $f_{\theta }(x)\to y$ or $p_{\theta }(y|x)$
  • Regression models
    • Linear Regression
    • Ridge Regression
    • Lasso Regression
    • Elastic Net
    • Support Vector Regression (SVR)
    • Random Forest Regression
    • Gradient Boosting Regression
    • Gaussian Process Regression
    • Neural Network Regression
  • Classification models
    • Logistic Regression
    • Softmax Regression
    • Naive Bayes
    • k-Nearest Neighbors (kNN)
    • Support Vector Machine (SVM)
    • Decision Tree
    • Random Forest
    • AdaBoost
    • Multilayer Perceptron (MLP)
    • CNN Classifier
    • RNN Classifier
    • Transformer Classifier

• Unsupervised Learning

  • Uses only input data to discover structure, distributions, clusters, low-dimensional representations, or anomalies.
  • Data form: $\mathcal{D}=\{x_i{\}}_{i=1}^n$
  • Clustering methods
    • K-Means
    • Gaussian Mixture Model (GMM)
    • Spectral Clustering
  • Dimensionality reduction
    • Principal Component Analysis (PCA)
  • Density estimation
    • Kernel Density Estimation (KDE)
    • Gaussian Mixture Model (GMM)
  • Anomaly Detection

• Self-supervised Learning

  • Builds supervision signals from data itself without manual labels.
  • Autoregressive prediction
    • n-gram Language Model
    • RNN Language Model
    • Transformer Language Model
    • GPT-style Next-token Prediction
  • Masked prediction
    • Masked Language Modeling (MLM)
    • BERT-style Pretraining
    • Masked Autoencoder (MAE)
    • Masked Image Modeling
  • Contrastive learning (to be refined)
    • Word2Vec
    • SimCLRbi
    • MoCo
    • BYOL
    • Barlow Twins
    • SwAV
    • CLIP-style Image-Text Contrastive Learning
  • Denoising objectives
    • Denoising Autoencoder
    • BART-style Denoising
    • Diffusion Denoising Objective
  • Predictive representation learning
    • Contrastive Predictive Coding (CPC)
    • Bootstrap Representation Learning

• Semi-supervised Learning

  • Uses a small labeled set and a large unlabeled set jointly.
  • Data form: ${\mathcal{D}}_L=\{(x_i,y_i)\},{\mathcal{D}}_U=\{x_j\}$
  • Typical methods
    • Pseudo-labeling
    • Consistency Regularization
    • FixMatch

• Reinforcement Learning

  • Learns policies from agent-environment interaction.
  • Policy form: $\pi (a|s)$
  • Objective: $\underset{\pi }{max}\mathbb{E}\left[\sum _{t=0}^T{\gamma }^t{r}_t\right]$
  • Core entities
    • State
    • Action
    • Reward
    • Policy
    • Value Function
    • Environment Model
  • Bandits
    • Multi-Armed Bandit
    • epsilon-Greedy
    • Upper Confidence Bound (UCB)
    • Thompson Sampling
  • Dynamic programming
    • Policy Evaluation
    • Policy Iteration
    • Value Iteration
  • Monte Carlo RL
    • Monte Carlo Prediction
    • Monte Carlo Control
  • Temporal Difference Learning
    • TD(0)
    • Q-Learning
    • SARSA
  • Deep reinforcement learning
    • Deep Q-Network (DQN)
    • Proximal Policy Optimization (PPO)
    • Soft Actor-Critic (SAC)
    • MuZero

• Imitation Learning (to be refined)

  • Learns policies from expert demonstrations.
  • Typical methods
    • Behavior Cloning
    • DAgger
    • Inverse Reinforcement Learning (IRL)
    • Maximum Entropy IRL
    • Generative Adversarial Imitation Learning (GAIL)

• Active Learning (to be refined)

  • Actively selects the most valuable samples for annotation.
  • Typical methods
    • Uncertainty Sampling
    • Query-by-Committee
    • Expected Model Change
    • Expected Error Reduction
    • Core-set Selection
    • Diversity-based Sampling

By Model Structure

Model structure describes how a model expresses functions, distributions, or policies.

• Linear Models: $f(x)=w^{\mathrm{\top }}x+b$
• Kernel Models: $K(x,x^{\prime })=\varphi (x{)}^{\mathrm{\top }}\varphi (x^{\prime })$
• Tree Models: recursive partitioning in feature space
• Ensemble Models: bagging, boosting, and model-combination families
• Probabilistic Graphical Models: Bayesian networks, MRFs, factor graphs, etc.
• Neural Networks: feedforward, CNN, RNN, Transformer, GNN, autoencoders

By Probabilistic Perspective

This view focuses on how models represent distributions, uncertainty, latent variables, and dependencies.

• Discriminative Models
• Generative Models
• Latent Variable Models
• Bayesian Models
• Energy-based Models
• Approximate Inference Methods

By Task Objective

Task objectives describe what the model is ultimately optimized to solve.

• Prediction / Classification / Regression
• Clustering / Dimensionality Reduction / Representation Learning
• Density Estimation / Generation / Structured Prediction
• Ranking / Anomaly Detection
• Decision Making / Planning
• Causal Inference