Playground
Hands-on experiments to build intuition. Each one runs in your browser and links to the lessons that explain the math.
How to use it #
Before you move a slider or press play, predict what will happen. Then check. A wrong prediction is the most useful moment: it shows exactly which idea to look at again in the linked lesson.
Experiments #
Neural network playground #
Train a small neural network in your browser. Change the data, the input features and the shape of the network, and see overfitting, depth versus width and feature engineering happen live. Tracks: Deep Learning, Optimization.
Optimizer race #
Race seven optimizers (gradient descent, momentum, Nesterov, AdaGrad, RMSprop, Adam, AdamW) on six classic loss surfaces. See why momentum helps in narrow valleys and why too large a step diverges. Tracks: Optimization.
Diffusion and flow matching in 2D #
Add noise to a 2D data set, see the score field that points back to the data, and sample with the reverse SDE, the probability-flow ODE and flow matching. Tracks: Generative Models, ML for Science.
Attention playground #
Build scaled dot-product attention by hand on a short sentence: drag the word vectors, edit the query, key and value matrices, and watch the attention weights and outputs change. Tracks: Deep Learning, Linear Algebra.
Tracking with the Kalman filter #
Follow a moving object through noisy sensor readings. See how the Kalman gain balances the model and the data, and compare with a particle filter. Tracks: Time Series, ML for Science.
Reinforcement learning in a gridworld #
Build a small world with walls, goals and traps. Solve it with value iteration and policy iteration, then watch a Q-learning agent learn it by trial and error. Tracks: Reinforcement Learning.
Gaussian process regression #
Place data points and watch a Gaussian process fit them. See the uncertainty grow far from the data and change the kernel to change the model's beliefs. Tracks: ML for Science, Probability and Statistics.
Clustering: k-means and Gaussian mixtures #
Watch k-means and the EM algorithm group points step by step, compare k-means++ with a random start, and see where k-means fails and EM helps. Tracks: Probability and Statistics.