AI & Machine Learning

From the mathematics of learning to frontier research and ethics.

This is the technical reference for machine learning on this site. It covers the statistical foundations of learning, the classical algorithms that still dominate tabular data, the theory and architectures of deep learning, the objectives models are trained on, how pretrained models are adapted and aligned, reinforcement learning, generative models, and the research and governance questions raised by frontier systems. The pages assume comfort with linear algebra, calculus, and probability; new to the subject? Start with AI Fundamentals (Simplified), which needs no math.

What the Terms Mean

Artificial intelligence is the broad field of building systems that perform tasks associated with human intelligence: perception, language, planning, and decision-making. Machine learning is the subfield in which that behavior is learned from data rather than programmed by hand. Deep learning is machine learning with many-layered neural networks, and foundation models — large networks pretrained on broad data and adapted to many tasks, including large language models (LLMs) and image generators — are the dominant form of deep learning today.

flowchart TB
    subgraph AI["Artificial intelligence"]
        direction TB
        SYM["Symbolic AI: search, logic,<br/>planning, expert systems"]
        subgraph ML["Machine learning"]
            direction TB
            CLS["Classical ML: linear models,<br/>trees and boosting, SVMs, clustering"]
            subgraph DL["Deep learning"]
                FM["Foundation models:<br/>LLMs, diffusion, multimodal"]
            end
        end
    end

Machine-learning problems are usually grouped by the kind of feedback the learner receives:

Paradigm Training signal Typical tasks Covered in
Supervised Labeled input–output pairs Classification, regression, forecasting Core ML Algorithms, Loss Functions
Unsupervised Unlabeled data only Clustering, dimensionality reduction, density estimation Core ML Algorithms, Generative Models
Self-supervised Labels derived from the data itself (next token, masked patch, added noise) LLM and diffusion pretraining, representation learning Loss Functions, Generative Models
Reinforcement Scalar reward from an environment or evaluator Games, robotics, RLHF, reasoning-model training Reinforcement Learning

Nearly all deployed systems are narrow: they perform well within the domain they were trained for. Modern foundation models are far more general than earlier systems — one model can write code, analyze images, and hold a conversation — but they remain unreliable outside their training distribution and on long, open-ended tasks. Whether and when systems will match humans across all cognitive work (artificial general intelligence, AGI) is contested; there is no agreed definition or test. The research and governance questions this raises are covered in Frontier Research & Ethics.

Pages in This Reference

Page Area What it covers
ML & Deep Learning Track Overview Hub for the four core pages below, with guidance on how they fit together
Machine Learning Foundations Foundations Statistical learning theory, bias–variance, optimization (SGD, Adam), kernels and SVMs, Gaussian processes, variational inference
Core ML Algorithms Foundations Linear and logistic regression, trees, random forests, gradient boosting, k-NN, clustering — the strongest baselines on tabular data
Deep Learning Theory Deep learning Universal approximation, backpropagation, initialization and normalization, the neural tangent kernel, double descent, generalization
Deep Learning Architectures Deep learning MLPs, CNNs, RNNs and LSTMs, attention and the Transformer, vision transformers, and post-Transformer sequence models
Loss Functions & Objectives Training Regression, classification, contrastive, ranking, distillation, language-model, preference, and generative objectives
Fine-Tuning & Transfer Learning Training Full and parameter-efficient fine-tuning (LoRA, QLoRA), instruction tuning, RLHF, DPO
Reinforcement Learning Training MDPs, value and policy methods, deep RL (DQN, PPO), and RL for language models
Generative Models Generation Diffusion and flow matching, GANs, VAEs, autoregressive models, discrete diffusion
Frontier Research & Ethics Frontier Scaling laws, reasoning models, interpretability, safety and alignment, fairness, privacy, regulation

Reading Order

The pages are self-contained, but each builds on the ones before it. The full path runs:

flowchart LR
    F["ML Foundations"] --> C["Core ML Algorithms"]
    C --> T["DL Theory"]
    T --> A["DL Architectures"]
    A --> L["Loss Functions"]
    L --> FT["Fine-Tuning"]
    FT --> RL["Reinforcement Learning"]
    RL --> G["Generative Models"]
    G --> FR["Frontier & Ethics"]

Shorter routes, depending on the goal:

Core Ideas

A handful of ideas recur on every page:

  • Learning is optimization. Training minimizes an objective by gradient descent, $\theta_{t+1} = \theta_t - \eta\,\nabla_\theta\mathcal{L}$; the architecture determines what functions are reachable, and the loss determines which one is found.
  • The objective encodes assumptions. Most losses are negative log-likelihoods under a noise model — MSE for Gaussian noise, cross-entropy for categorical labels — so choosing a loss is choosing what you believe about the data.
  • Generalization, not fit, is the goal. Performance on held-out data is what matters. Overparameterized networks generalize far better than classical theory predicted, which is still only partly understood.
  • Self-supervision at scale. Predicting the next token or the added noise turns unlabeled data into a training signal; pretraining on that signal, followed by fine-tuning and preference optimization, is how modern foundation models are built.
  • Scaling is predictable, with caveats. Pretraining loss follows power laws in parameters, data, and compute, and since 2024 inference-time reasoning has become a second scaling axis. Downstream capabilities scale less smoothly than loss.
  • Capability and responsibility grow together. Evaluation, interpretability, fairness, privacy, and security are engineering requirements, and increasingly legal ones.

See Also