#live-training — Supervised learning

Six algorithms learning before your eyes, like a video: REC, timecode, subtitles, live metrics. Watching is free; touching the model is Premium.

What you'll play with

Welcome to #live-training 🎬. Here, for once, you don't set anything up first: you watch. The scene plays a film in six episodes, one per algorithm: a line learning to predict a price, a boundary separating pass from fail, a map of neighbors drawing itself row by row, a tree carving up the plane, a forest casting votes, and a margin growing wider. REC dot, timecode, subtitles, live metrics: each episode lasts under a minute and the next one follows on. Watching is free, on a loop, as much as you like. Touching the model — learning rate, k, depth, margin, predicting a point — is Premium: if you try, I'll tell you with an "Oops".
The film is rolling. Pause it whenever you like: /pause. The REC dot gives way to PAUSED, the timecode freezes, and you can read the metrics at leisure.
Speed up: /speed 2. The number of frames per second doubles. The slow episodes (the tree makes one split every 0.8 s) turn snappy; the gradient descents fly by.
Jump straight to an episode: /episode knn. The film restarts at the beginning of episode 3: k-nearest neighbors has nothing to train, so we watch the decision map being computed pixel by pixel, then an unknown flower looking for its neighbors.
Missed a detail? /replay restarts the current episode from the first frame. Same seed, same draw, same film: it is deterministic on purpose — to compare, you need to watch the same thing again.
Your turn. Viewer side (free): /episode tree, /episode forest, /episode svm, /speed 0.5, /subtitles to hide or restore the subtitles, /autoplay to loop or stop at the end of an episode. Director side (Premium): /lr 0.05 a cautious descent, /lr 1 a descent that goes off the rails, /k 1 nervous boundaries, /depth 6 a tree that overfits, /trees 40 a dense forest, /margin 20 an uncompromising margin, /noise 0.9 mixed-up data, /seed 12 another draw, /predict 0.7 0.3 to drop a star and read the model's verdict. Oops… those need the subscription — I'll tell you how. Next: #linear-regression to manipulate the line in 3D, or #knn to tune k yourself.

Channel commands

/episode <linear|logistic|knn|tree|forest|svm|next> — Jumps to an episode (or the next one) and plays it from the start.
/pause — Pauses the film (the timecode freezes).
/play — Resumes playback of the film.
/speed <0.5|1|2|4> — Playback speed: 0.5 (slow motion) to 4 (fast forward).
/replay — Restarts the current episode from the first frame (same draw).
/subtitles — Shows or hides the teaching subtitles.
/autoplay — Turns automatic episode chaining on or off.
/reset — Returns to the first episode and to the default settings.
/lr <0.01..1> — 🔒 Premium · Learning rate of the gradient descents (episodes 1, 2, 6).
/noise <0..1> — 🔒 Premium · Spread of the data: 0 = crisp clouds, 1 = heavily mixed.
/points <n=20..200> — 🔒 Premium · Number of points (one in four is held out for testing).
/seed <1..99> — 🔒 Premium · Another random draw of the data (same settings).
/k <1..15> — 🔒 Premium · Number of neighbors for KNN (episode 3).
/depth <1..6> — 🔒 Premium · Maximum depth of the decision tree (episode 4).
/trees <1..40> — 🔒 Premium · Number of trees in the random forest (episode 5).
/margin <C=0.1..20> — 🔒 Premium · SVM constant C: penalty for points inside the margin (episode 6).
/predict <x=0..1> <y=0..1> — 🔒 Premium · Drops a star at (x, y) and reads the trained model's answer.

Glossary

Epoch: One full pass over the training data, with one update of the parameters. In the film, every frame of episodes 1, 2 and 6 is an epoch: the line or the boundary moves a little at each one.
Loss function: The number training tries to drive down: MSE for regression, log-loss for logistic regression, hinge loss for the SVM. The curve going down to the right of the film is that number.
Convergence: The moment when updates become negligible: the loss barely moves anymore, the model has "arrived". The last subtitles of the gradient-descent episodes point it out.
Divergence: When the learning rate is too large, every step overshoots the bottom of the bowl and climbs higher than the previous one: the loss explodes. The film stops on a red line and an "Oops".
Decision boundary: The line (or surface) where the model changes its mind between two classes. Straight for logistic regression and the linear SVM, jagged for KNN, a staircase for a tree, blurry then crisp for a forest.
Test set: Points set aside (the hollow circles) that the model never sees during training. Honest accuracy is measured on them: classifying correctly what you learned by heart proves nothing.
Support vector: A training point sitting on the margin or inside it (ringed in yellow in episode 6). These are the only points that determine an SVM's boundary: removing the others would not move it.
Gini index: A measure of impurity of a group of points: 0 if a single class, maximal if the classes are in equal shares. At every node, a decision tree picks the split that lowers the average impurity of the two halves the most.
Bootstrap and bagging: Drawing n points with replacement from the n original ones gives a slightly different sample; training one model per sample and then letting the ensemble vote is bagging. A random forest is bagging of trees.

Other channels in Supervised learning

#live-training — Six algorithms learning before your eyes, like a video: REC, timecode, subtitles, live metrics. Watching is free; touching the model is Premium.
#linear-regression — Fit a line: least squares, residuals, MSE, R² and gradient descent — the first brick of every supervised model.
#logistic-regression — Classify into two categories: sigmoid, decision boundary, threshold and log-loss — and why a line is not always enough.
#decision-trees — A tree that carves the plane into rectangles: Gini, entropy, depth, pruning — and the overfitting you can see with your own eyes.
#knn — k nearest neighbours: classify by resemblance, pick k, change the distance — and watch the boundary smooth out or shatter.
#svm-margins — Support vector machines: the widest possible margin, the C parameter, and the RBF kernel that curves the boundary.
#classification-metrics — Precision, recall, F1, confusion matrix, ROC and AUC: reading a classifier honestly, especially when classes are imbalanced.