ML & AI Learning Hub

A research-grade walkthrough of the modern machine-learning and AI stack, from the perceptron to frontier research problems. Every page follows the same structure: prerequisites, learning outcomes, a TL;DR card, an intuition section, maths deep-dives, interactive demos or worked code, comparisons, pitfalls, interview Q&A, a self-check quiz, and verified references. The site is built for two audiences: people learning the field for the first time, and engineers brushing up before a research-grade interview.

Learning path

The Foundations sequence is meant to be read in order. Each page assumes the previous ones.

1. History of AI & ML - timeline from McCulloch-Pitts (1943) to GPT-4 (2023), with the two AI winters explained.
2. Why ML differs from software engineering - opinionated framing of the three axes on which ML breaks traditional software workflows.
3. AI / ML / DL / GenAI map - the conceptual hierarchy that places every technique inside a single Venn diagram.
4. Math you need - linear algebra, calculus, probability, statistics, and information theory at the depth required to read modern papers.
5. The ML workflow - end-to-end lifecycle from problem framing through monitoring and retraining.
6. Data & leakage - split strategies, leakage patterns, and three real-world horror stories.
7. Bias, variance, over/under-fitting - the classical decomposition plus the modern double-descent picture.
8. Evaluation metrics - classification, regression, and ranking metrics with worked-out tradeoffs.
9. Perceptron - the canonical reference page: model, learning rule, Novikoff convergence theorem, XOR limitation, modern descendants.

Reinforcement Learning

Reinforcement learning bridges the gap between supervised learning and autonomous decision-making. Agents learn by interacting with an environment, receiving rewards, and updating policies to maximise long-run return.

• Foundations of RL - MDPs, Bellman equations, value iteration, Q-learning, DQN, policy gradients (REINFORCE), PPO, SAC, and AlphaGo/AlphaZero lineage.

Modern AI

Ten pages covering the architectures and training recipes that define today’s state of the art, from attention mechanisms to multimodal LLMs.

• Attention Mechanisms - Bahdanau, scaled dot-product, multi-head, FlashAttention.
• Transformers - architecture, positional encoding, KV cache, scaling.
• Tokenization & Embeddings - BPE, WordPiece, SentencePiece, embedding spaces.
• Encoder, Decoder & Encoder-Decoder Models - BERT, GPT, T5 families.
• Scaling Laws & Training - Chinchilla, compute-optimal training, emergent abilities.
• Fine-tuning & PEFT - LoRA, QLoRA, adapter methods, RLHF, DPO.
• Retrieval-Augmented Generation - BM25, DPR, ColBERT, chunking, RAGAS evaluation.
• Agents & Tool Use - ReAct, function calling, planning, evaluation.
• Diffusion Models - DDPM, score matching, DDIM, latent diffusion, CFG.
• Multimodal Models - CLIP, ALIGN, BLIP-2, Flamingo, LLaVA, Gemini.

Production & MLOps

Engineering practices for deploying, serving, and monitoring ML systems at scale.

• LLM Serving & Inference - KV-cache memory math, PagedAttention (vLLM), continuous batching, speculative decoding, GPTQ, AWQ, throughput vs latency.
• MLOps - experiment tracking, feature stores, CI/CD for ML, the four monitoring axes, shadow deployment, canarying, A/B test design.
• Evaluation & Benchmarks - static benchmarks vs LLM-as-judge, contamination, LMSYS Arena, HELM, BIG-bench, SWE-bench, GAIA.

Responsible AI

Fairness, alignment, privacy, and safety for production ML systems.

• Safety & Alignment - RLHF, Bradley-Terry derivation, DPO closed-form, Constitutional AI, reward hacking, jailbreaks, prompt injection.
• Fairness & Explainability - fairness impossibility theorem, SHAP, LIME, Integrated Gradients, DP-SGD, federated learning.

Research Frontiers

• Open Problems in AI - hallucination, reasoning (chain-of-thought, o1, DeepSeek-R1), world models, embodied AI, mechanistic interpretability, Mamba, RWKV, RingAttention, data quality.

How to use this site

Classical ML

Eleven pages covering the workhorses of supervised, unsupervised, and recommendation problems: linear and logistic regression, k-NN, decision trees, random forests, gradient boosting, SVMs, naive Bayes, clustering (K-Means, DBSCAN, GMM), dimensionality reduction (PCA, t-SNE, UMAP), and recommender systems. Each page derives the model from first principles, contrasts it with adjacent algorithms, and ships an interactive demo or worked code so you can build intuition before the maths hardens.

Start with Linear Regression for the canonical least-squares treatment, or jump to Gradient Boosting to see the algorithm that still wins most tabular Kaggle competitions.

Deep Learning

Eight pages tracing the path from a single hidden layer to modern generative models: MLPs, backpropagation and optimisers, regularisation in deep nets, CNNs, RNNs/LSTMs/GRUs, autoencoders and VAEs, GANs, and a capstone of practical end-to-end projects. The maths is rigorous (chain rule, vanishing-gradient analysis, ELBO derivation) and the demos run in the browser via lazy Pyodide, so nothing requires a local Python install.

Start with the Multi-Layer Perceptron for the simplest universal approximator, or jump to Convolutional Neural Networks for the architecture that put deep learning on the map.

Modern AI

Ten pages on the post-2017 stack that produced ChatGPT, Stable Diffusion, and the agent ecosystem. The thread runs: scaled dot-product attention, the full Transformer block, tokenization and embeddings, the encoder / decoder / encoder-decoder split, the scaling laws that justified the spend, parameter-efficient fine-tuning, retrieval-augmented generation, agents and tool use, diffusion models, and multimodal foundation models. Every page derives the load-bearing equations (softmax attention, RoPE, the DDPM ELBO and epsilon-prediction objective, the InfoNCE contrastive loss), shows the architecture at the level you would draw it on a whiteboard, and ends with research-grade interview questions.

• Attention mechanisms - scaled dot-product, multi-head, masking, and why $\sqrt{d_k}$.
• The Transformer architecture - encoder, decoder, pre-norm vs post-norm, and the KV-cache memory math.
• Tokenization & embeddings - BPE, WordPiece, SentencePiece, and the geometry of learned vectors.
• Encoder, decoder & encoder-decoder models - BERT vs GPT vs T5 and when each pattern wins.
• Scaling laws & training - Kaplan, Chinchilla, compute-optimal frontiers, and the trillion-token playbook.
• Fine-tuning & PEFT - SFT, LoRA, QLoRA, adapters, prefix tuning, and DPO.
• Retrieval-augmented generation - dense retrievers, reranking, hybrid search, evaluation, and failure modes.
• Agents & tool use - ReAct, function calling, planning, multi-agent systems, and benchmarks.
• Diffusion models - DDPM ELBO, score matching, DDIM, classifier-free guidance, and latent diffusion.
• Multimodal models - CLIP, BLIP-2, Flamingo, LLaVA, and the architectural patterns of vision-language fusion.

Start with Attention mechanisms for the equation behind everything, or jump to Diffusion models for the cleanest derivation on the site.