Explore
101 visualisers you can step through forwards and backwards, with synced code and a reason for every move.
Showing 25 of 101
Pastel = track · chip = your mastery
- Linear AlgebraLinear TransformationsSee a 2×2 matrix move the whole plane: its columns are where î and ĵ land, and its determinant scales area.
- Linear AlgebraEigenvectorsSweep a vector around the circle to find the directions a matrix only stretches: A·v = λ·v.
- Linear AlgebraSVD Image CompressionRebuild a picture from its singular value decomposition one rank-1 piece at a time, and see how few pieces most images need.
- CalculusDerivativesShrink a secant line until it becomes the tangent, then trace the slope at every point to draw f′.
- CalculusThe Chain Rule (Backpropagation)Run a computation graph forward, then push gradients back node by node: local derivative times upstream gradient.
- OptimizationGradient DescentWalk downhill on a contour map by stepping against the gradient, and see what the step size does.
- OptimizationMomentum and AdamRace gradient descent, momentum, RMSProp and Adam on the same surface, from the same start.
- ProbabilityBayes' RuleUpdate a belief with evidence by drawing everyone as a square: the answer is the share of the evidence strip where the hypothesis holds.
- ProbabilityCentral Limit TheoremAverage samples from any population, flat, skewed or split in two, and watch the averages pile up into a bell curve whose spread shrinks like σ/√n.
- Information TheoryEntropy and KL DivergenceMeasure information as surprise in bits, then see what believing the wrong distribution costs: cross-entropy and KL divergence.
- Numerical MethodsFloating-Point NumbersSee how a float32 packs a sign, an exponent and 23 fraction bits, why 0.1 is not stored exactly, and how far apart neighbouring floats are.
- Machine LearningLinear RegressionFit a line by gradient descent on the mean squared error, with every residual drawn as a real square, then compare with the exact least-squares line. Drag the points to change the data.
- Machine LearningLogistic RegressionTrain a single sigmoid neuron on the cross-entropy: watch the probability regions sharpen and the boundary swing into place, and see why it can't separate XOR or circles.
- Machine LearningK-Means ClusteringGroup points by alternating two steps, joining the nearest centroid and moving centroids to their means, and see where k-means finds the right clusters and where it gets stuck.
- Machine LearningPerceptronWatch a single neuron learn a line by nudging it towards every point it gets wrong, and see why it stops on separable data but never on XOR.
- Machine LearningDecision TreesGrow a decision tree one question at a time, choosing each split to make the regions as pure as possible, and see where deep trees start to overfit.
- Machine LearningGradient BoostingBuild a strong model from tiny trees: each round fits a one-split stump to what the model still gets wrong and adds a shrunk copy of it.
- Machine LearningNeural Network PlaygroundTrain a small neural network on XOR, circles or moons and watch hidden layers bend a straight decision boundary into a curve, with every weight drawn as it changes.
- Deep LearningConvolutionSlide a 3×3 kernel over a picture to build a feature map, then apply ReLU and max-pooling: how a CNN finds edges.
- Language ModelsByte-Pair EncodingLearn a tokenizer from text: glue the most frequent pair of neighbours into a new token, merge after merge, then split a word it never saw.
- Language ModelsAttentionWatch a query score every key, softmax turn the scores into weights, and the output blend the values; drag the query to steer it.
- Language ModelsSampling and TemperatureTurn a language model’s logits into one next word: softmax with temperature, top-k and top-p cuts, and a random draw.
- Language ModelsKV CacheGenerate text token by token and see why caching keys and values turns each step from recomputing everything into one new row.
- Reinforcement LearningQ-LearningAn agent learns a gridworld from rewards alone: watch Q-values spread back from the goal, the greedy arrows settle, and the discount decide between a near and a far reward.
- Generative ModelsDiffusion ModelsNoise a cloud of points into a blob, then watch a trained denoiser turn fresh noise back into the shape, one small step at a time.