Fig. 00 / How the domains connect223 nodes13 domains36 systems
ML Atlas / Index of machine learning
A map of machine learning, built to be explored.
Every model, technique and concept is a place on the map. See where it fits, touch it in a live demo, and meet the real systems that use it, from Random Forests to Claude.
- Topics
- 223
- Domains
- 13
- Live demos
- 13
01 / Index
Domains
- 01FoundationsData, loss, gradients and the ideas every model shares.19 topics
- 02Learning ParadigmsThe different ways a machine can learn: with labels, without, or by trial and error.10 topics
- 03Classical Machine LearningRegression, trees, ensembles, clustering: the workhorses of tabular data.40 topics
- 04Neural Network FundamentalsNeurons, layers, backpropagation and the tricks that make deep nets train.15 topics
- 05Deep Learning ArchitecturesCNNs, RNNs, attention and Transformers: how networks are wired.14 topics
- 06Generative AIModels that create: autoencoders, GANs and diffusion.18 topics
- 07Language and LLMsFrom tokens to chat assistants: how large language models work.36 topics
- 08Computer VisionTeaching machines to see: classification, detection, segmentation.16 topics
- 09Speech and AudioSpectrograms, speech recognition and synthetic voices.9 topics
- 10Multimodal AIModels that connect text, images and sound.5 topics
- 11Reinforcement LearningAgents that learn from rewards by trial and error.16 topics
- 12ML Engineering and MLOpsGetting models into production and keeping them healthy.12 topics
- 13Responsible AI, Safety and SecurityFairness, interpretability, attacks and alignment.13 topics
02 / Start here
ML in 60 minutes
A first path through the essentials, in order.
03 / Learn by touching
Live demos
Real models and exact math, running in your browser. Change something and watch what happens.
- Demo 01 / k-Means Clusteringk-means, one step at a timek-means only ever alternates two simple moves, and it always settles, but where it settles depends on where the centroids start.
- Demo 02 / GPTHow a language model picks the next wordThe model only ever outputs probabilities. Temperature, top-k and top-p decide how adventurous the pick is, and the same model can sound careful or chaotic.
- Demo 03 / AttentionWhere does each word look?Each attention head is a different way of reading the sentence. Some just look at the previous word; one learned, without being told, that “it” usually means the animal.
- Demo 04 / BackpropagationBackpropagation, one multiplication at a timeBackprop is just the chain rule applied backward: each node multiplies the gradient it receives by its own local slope and passes it on.
- Demo 05 / Gradient DescentRoll downhill: gradient descent, momentum and AdamSame landscape, same starting point, different update rules. The learning rate decides whether you crawl, glide or fly off the map.
- Demo 06 / Classifier-Free GuidanceFrom noise to data: diffusion, exactlyDestroying data with noise is easy and needs no learning. A diffusion model learns to undo it one small step at a time, and guidance pushes each step harder toward what you asked for.
- Demo 07 / Object DetectionFrom hundreds of guesses to a few boxesDetectors guess many overlapping boxes. A confidence threshold and non-max suppression (using IoU) clean them up, and the two thresholds trade missed objects against false alarms.
- Demo 08 / Q-LearningQ-learning in a gridworldNobody tells the agent the way. Reward at the goal slowly spreads backward through the Q-values, one step per episode, until a path emerges.
- Demo 09 / Audio ClassificationSee what sound looks likeA spectrogram turns sound into a picture of pitch over time, which is what speech and audio models actually read. Window size trades sharp timing against sharp pitch.
- Demo 10 / Bias and FairnessHow biased data becomes biased decisionsA model faithfully learns the unfairness in its training labels. Removing the sensitive column is not enough when other features act as proxies; fairness has to be measured and chosen.
- Demo 11 / Semi-Supervised LearningSame data, three ways to learnWith only a few labels, supervised learning draws a crude boundary. Semi-supervised learning lets the shape of the unlabeled data spread the labels, which works brilliantly when clusters match classes and fails when they do not.
- Demo 12 / CLIPMatching pictures and words with CLIPCLIP puts images and captions in one embedding space, so it can label images and search them with any words, including mistaking written text for the thing it names.
- Demo 13 / PerplexityShrinking a model by rounding its weightsFewer bits per weight make models smaller and cheaper. Done naively, a few outlier weights wreck the model; giving each small group its own scale keeps GPT-2 almost intact down to 4 bits.