Neural Network
A neural network as a function: every neuron visible, ReLU kinks as building blocks of the output, weights editable per neuron, parameter count; curve, plane and digits modes.
Neural networksTry itCollection
From neurons and gradients to transformers, RLHF and agents: machine learning you can touch.
Apps
A neural network as a function: every neuron visible, ReLU kinks as building blocks of the output, weights editable per neuron, parameter count; curve, plane and digits modes.
Neural networksTry itLogits become a distribution via softmax(z/T): temperature, greedy, sampling, top-k and top-p with a visible cut-off line, entropy, a decision tree of the generation and loops at T = 0.
Neural networksTry itText splits into tokens: byte-level BPE is trained live, the merge slider shows frequent words fusing into one token, plus token details, a language comparison and the share of the context window.
Linguistics & phoneticsTry itText from counts of the last n − 1 tokens: count table P(w|h), the path from gibberish to plausible text to verbatim copying, the explosion of possible contexts, smoothing and backoff; German and English corpora.
Linguistics & phoneticsTry itA model is a form with parameters: polynomials of degree 0 to 12 by least squares, residuals as squares (area = loss), held-out error over the degree and parameters that explode when overfitting.
Probability & statisticsTry itOne artificial neuron: weighted sum, activation, a decision line in the input plane with the w arrow and bias offset; perceptron rule and gradient step solve AND and OR, XOR stays at 75 %.
Neural networksTry itA drawn digit flows through a real network trained on MNIST (784 → 64 → 32 → 10): weight images per neuron, forward wave, softmax bars and test digits it gets confidently wrong.
Neural networksTry itTraining lab for neural networks in 3D: neurons as tiles showing their output, edges by weight, spiral, XOR, moons, layers, activations and regularisation live; plus a digits mode.
Neural networksTry itThe gradient (f_x, f_y) lies in the plane, is perpendicular to the contour line and points uphill where it is steepest. A rotatable direction vector u gives the directional derivative ∇f · u: largest along the gradient, zero along the contour line.
Analysis (multivariable)Try itLoss surface of a real line fit in 3D: θ ← θ − η∇L with stability limit 2/λmax, two valleys and a saddle, GD, momentum and Adam racing, SGD noise and the fitted line in the model panel.
Neural networksTry itOne whole learning round on a real network (64 → 8 → 8 → 10, 682 parameters): it takes a sloppy 3 for an 8. Error, loss as one square, slope per weight, blame layer by layer backwards, a ball rolls downhill over a real loss surface, and afterwards it recognises the 3. Draw yourself, change the learning rate, practise a thousand times.
Neural networksTry itBackpropagation on a 2-2-2 network with real numbers in six steps: forward pass, loss, δ at the output and in the hidden layer, gradient per weight, only the update lowers the loss — checked numerically.
Neural networksTry itReading training dynamics from the loss curves: learning rate, batch size, epochs, training and validation loss, early stopping at the validation minimum, L2, dropout, augmentation and more data against overfitting.
Neural networksTry itReal word vectors — English GloVe, German fastText (cc.de.300): relations as parallel arrows, king − man + woman ≈ queen with hits, near misses and learned bias, cosine against distance and sentences as a path with next-word candidates from GPT-2 or GerPT2.
Linguistics & phoneticsTry itWords as vectors in a space of meaning: similarity as cosine in the full space, analogy arrows, neighbours and three 3D shadows (map, PCA, axes) with how much neighbourhood each one keeps.
Neural networksTry itStatic vs. contextual word vector in 3D: an ambiguous word starts between two islands of meaning and is pulled layer by layer by the context words, in proportion to their attention weights.
Linguistics & phoneticsTry itSelf-attention made visible: query, keys, softmax(QKᵀ/√d)·V, five specialised heads, a matrix heat map with causal mask, a second layer and a by-hand calculation with three tokens.
Neural networksTry itThe GPT-2 architecture to walk through: residual stream through 12 blocks, masked attention as bridges, MLP d → 4d → d, unembedding and softmax, one token at a time; parameters per component and GPT-3 for comparison.
Neural networksTry itA small GPT runs in the browser, every number visible: tokens, embedding, Q·Kᵀ and the attention matrix with mask, residual stream, MLP and probabilities with temperature and top-k/top-p; zooms for every computation step.
Neural networksTry itYou compare answers: a reward model learns from them (σ(r_A − r_B)), the policy shifts under a KL leash; with a loose leash the flattering answer wins (reward hacking); plus the DPO route.
Neural networksTry itRetrieval-augmented generation step by step: chunks, embeddings, search by cosine, BM25 or hybrid, top-k cut-off, prompt with token budget and an answer with sources; missed chunks lead to hallucination.
Neural networksTry itElementary cellular automata with all 256 rules: rule 90 jumps to any generation by formula, rule 30 must compute every step (irreducibility); plus a prediction game and an LLM's compute budget per token.
Computer scienceTry itLeast-squares line on lecture data: errors as squares, SST = SSR + SSE as a right triangle, SSE as a bowl over slope and position; Anscombe, a parabola and a confounder show what R² does not reveal.
Probability & statisticsTry itMany training sets from the same world as a family of models: average model against the truth (bias), spread of the models (variance) and the exact decomposition test error = bias² + variance + noise over polynomial degree or ridge λ.
StatisticsTry itA query point, its k neighbours with distance shape and majority vote: decision regions from k = 1 (jagged) to k = 40 (smooth), train and test accuracy over k, metric and feature scaling change the neighbourhood.
Machine LearningTry itA bank learns from 150 simulated past customers (age and net income) who pays back a loan. Three learning methods draw the boundary: a straight line misses the U shape in age (underfitting), a curve with age² fits, the nearest neighbour gives every customer an island — 0 % wrong in training, but clearly worse on 150 new customers (overfitting). Drag Max Müller's application across the map.
Machine LearningTry itTwo features, a linear score, a sigmoid surface in 3D: the threshold plane cuts it in a straight line, log loss per point, gradient descent step by step, XOR as the model's limit.
StatisticsTry itPrior times per-feature likelihoods: posterior field with marginal densities, axis-aligned ellipses and a quadratic boundary; correlated data show what the independence assumption misses; plus a spam mode with log-odds.
Machine LearningTry itMaximum margin and support vectors, solved exactly with SMO: points without a ring do not move the boundary, soft margin C, slack ξ, linear, lift (paraboloid in 3D) and RBF kernels.
Machine LearningTry itAxis-parallel splits by information gain: ΔI for every candidate threshold, tree diagram linked to the plane, depth, minimum leaf size and pruning, train against test accuracy over the depth.
Machine LearningTry itEight classifiers learn on the same data: kNN, logistic regression, linear and RBF SVM, naive Bayes, decision tree, random forest and a neural network — their decision boundaries reveal the model family, train and test accuracy side by side expose overfitting.
Machine LearningTry itSplit, validation curve, k-fold cross-validation and a one-time test on a polynomial regression — plus data leakage and tuning on the test set as counterexamples.
Machine LearningTry itSlide a classifier's threshold and watch cases move between TP, FP, FN and TN — accuracy, precision, recall, specificity and F1 with their formulas, the precision–recall curve and the accuracy paradox with rare positives.
Machine LearningTry itTwo score distributions, the threshold sweeps: the ROC curve is traced point by point, AUC as the share of correctly ranked pairs, ROC stays the same at any prevalence, two models with crossing curves compared.
Machine LearningTry itMany random decision trees vote: the majority smooths the jagged boundaries of single trees and lowers the test error — bootstrap samples, feature randomness, out-of-bag error and test error over the number of trees.
Machine LearningTry itDecision stumps trained one after another on weighted data: mistakes grow, the next stump targets them, vote weight α per round, the ensemble's exact staircase boundary, train and test error over up to 60 rounds.
Machine LearningTry itLloyd's algorithm as a chain of states without labels: assign and move centres, Voronoi cells, inertia per half-step, random vs. k-means++, local minima, elbow with silhouette and typical failure cases.
Machine LearningTry itPrincipal axes of a 3D point cloud as eigenvectors of the covariance: projection onto k components with perpendiculars, reconstruction error = sum of the discarded λ, scree plot, standardising and the difference from the regression plane.
Machine LearningTry itThe same prompt through pretraining, SFT, RLHF/DPO and LoRA: the knowledge comes from pretraining, the later stages mainly shift which answer is likely.
Neural networksTry itAn LLM agent is a loop: the model writes tool calls as text, a program executes them and every observation grows into the context — with and without tools compared, including a tool failure and a prompt injection.
Neural networksTry itA model trained on past hiring decisions sorts 200 applications. Delete the group from the data and the ratio barely improves, because a proxy feature carries the bias on.
AI literacyTry itA random number generator is a fixed rule: same seed, same sequence, and eventually the loop closes. Middle-square, a linear congruential generator with freely adjustable parameters, RANDU, a shift register and xorshift side by side — good ones and deliberately bad ones. Four views reveal what a list of numbers hides: the rule with values plugged in, the cycle as a rho shape, the scatter of consecutive pairs plus a rotatable cloud of triples (RANDU collapses into 15 planes), and a histogram.
CryptographyTry itAn AI answer sounds certain and cites sources. Break it into claims and check each against the original document: supported, refuted or made up.
AI literacyTry itWhether a number is even is told by a rule: the last digit. Whether a pixel image is a handwritten seven is told by none: a real pixel rule gets 6 of 18 images wrong. Classical programming means rules plus data give answers, machine learning means data plus answers give rules. A real logistic regression turns its mixing desk of 8 knobs on 4800 examples, the error drops from 50 to 5.8 percent, then it recognises new sevens.
Machine LearningTry itFace detection as in the 1990s: an eye mask slides over a drawn street scene by cross-correlation and finds the frontal face, but misses shadow, sign and profile and takes headlights for eyes (1 of 5). A small neural network trained on 3909 drawn examples with answers finds all 5 faces in the same scene without a false alarm.
Machine LearningTry itFrom the table into space: 24 flats with m², rooms and area fly as points into a 3D feature space, the colour is the label (price or “sold within a month”). For a new flat “?” the k nearest neighbours answer — as regression with a price, as classification with yes or no. Row = example, column = feature, last column = label.
Machine LearningTry itThousands of particles flow step by step from Gaussian noise into shapes: forward and reverse process, DDPM vs. DDIM and classifier-free guidance. An image panel shows the same in pixel space, coarse shape first and details last.
Neural networksTry itA short-video feed as a model: watch time, likes and rewatches become a signal, the interest profile learns, candidates are ranked. With an exploration slider you can see how a feed that only exploits narrows to one topic, and how searching, resetting and exploring widen it again.
Film & mediaTry itCollections
Mathematics with studded bricks, built instead of drawn: how many buildings six bricks make, fractions as parts of a brick, little Gauss as a staircase and gears that only meet again after the lowest common multiple.
4 appsOpen collectionInstruments from the inside: the grand piano action, synthesizer, electric guitar, electric bass and drum kit, all of them together as a band, bowed strings on violin and cello, organ, trumpet, flute, saxophone, handpan, guitar and keyboard. Cut them open, play them yourself and hear how the sound is made.
16 appsOpen collectionSentence structure, participles, the accusative with infinitive, verb forms, sequence of tenses, small words, rhetorical figures and the hexameter in real passages from Seneca and Ovid.
8 appsOpen collectionGalvanic cells, reaction mechanisms, equilibrium and the Haber-Bosch process, acids and bases, dyes, plastics, intermolecular forces and molecular structure to try out.
18 appsOpen collectionPhotoelectric effect, potential well, Franck-Hertz, Millikan, diffraction gratings, standing waves, particles in fields, induction and oscillations to try out.
23 appsOpen collectionGene regulation, PCR, genetic crosses, protein biosynthesis, evolution, cellular respiration, hormones and ecosystems to try out.
24 appsOpen collectionPlanning and steering projects: the iron triangle, stakeholders, network diagrams, earned value, schedule risk, team size, Kanban and waterfall against Scrum.
8 appsOpen collectionFrom sequences to loans: arithmetic and geometric growth, limits and series, savings plans, annuities and outstanding balance, compounding within the year and perpetuities.
14 appsOpen collectionReading the traces: burial mounds in section, radiocarbon dating, settlements in the landscape and the tomb of Tutankhamun.
5 appsOpen collectionBuild it yourself like with a construction kit: gears, ratios, idler and worm, plus differential, watch movement and levers.
9 appsOpen collectionWhy robot arms are heavy, how six axes guide a hand and how control holds the load.
6 appsOpen collectionTables, keys, joins, groups and subqueries you can touch: how a database really reads.
7 appsOpen collectionGenerative AI in strong images: what it can do, how it writes, why it convinces, how to use it well and where to take care.
7 appsOpen collectionEngine, brakes, high voltage, CAN bus and chassis: technology and diagnostics of the modern vehicle.
9 appsOpen collectionHeat in the city centre: heat islands, shade, trees and unsealing, using the old town of Reutlingen as an example.
7 appsOpen collectionReading the small print, spotting subscription traps, placing risk figures and making texts understandable.
5 appsOpen collectionFrom equation to curve: initial values, linear equations, substitution and the worlds in which they apply.
7 appsOpen collectionRules and physics in sport: offside in football, spin and the Magnus effect, the walls in padel.
3 appsOpen collectionCalculus in the mountains: surfaces and contour lines, partial derivatives as slices, the tangent plane, the gradient, peaks, valleys and passes, and the optimum under a constraint.
10 appsOpen collectionFrom area to the integral: Riemann sums, antiderivatives, integration rules, units and improper integrals.
5 appsOpen collectionMean, median, mode, variance, box plot and levels of measurement: descriptive statistics to try out.
12 appsOpen collectionFrom coordinates to the inverse: vectors, dot and cross products, linear equations and matrices in pictures.
14 appsOpen collectionProduction principles, waste, one-piece flow, value streams, bills of materials and routing: how a factory is organised.
11 appsOpen collectionWhat the CRA requires of manufacturers: checking whether it applies, meeting reporting deadlines, keeping software bills of materials.
5 appsOpen collectionReducing, common denominators, multiplying, dividing and mixed numbers: fractions step by step.
13 appsOpen collectionProofs of special beauty to touch: Zagier's windmills, the museum guards and Buffon's needle.
3 appsOpen collectionHow oscillations break down: coefficients, symmetry, phasor chains, the Gibbs phenomenon and the plucked string.
7 appsOpen collectionWhat happens inside the machine: objects, methods, operators, slicing and sorting in Python, visible step by step.
12 appsOpen collectionFrom the Patent Motor Car to the Z3: famous machines and experiments cut open in 3D.
12 appsOpen collectionKeyboard, intervals, scales, chords, the circle of fifths and groove: music theory you hear and play.
9 appsOpen collectionHow sound is made and shaped: synthesizers and samplers, envelopes and overtones, waveforms and timbre, voice and orchestra, room acoustics and sound scenography — all audible in your browser.
26 appsOpen collectionStaging space, sight, light, sound and paths: visitor flows, showcase lighting, room acoustics, sound islands, projection mapping and the pavilions of world expos to try out.
17 appsOpen collectionMathematical questions open for decades, for which an AI model proposed solutions in 2026, as interactive apps: understand the problem, experience the result.
32 appsOpen collectionDNA, replication, transcription, translation and cell division as explorable 3D models.
27 appsOpen collectionPerspective, ornament, structure and space: how art and architecture work with mathematics.
25 appsOpen collectionFrom the project plan via data basis, ABC and KPIs to high-bay warehouse, order picking, value stream, factory layout, risks and FMEA: warehouses and factories as fully calculated 3D models.
10 appsOpen collectionHyperloop, autonomous vehicles, warehouses and conveyors: logistics as interactive simulation.
17 appsOpen collectionHeart, eye, voice, skeleton and nerves: the human body in 3D, layer by layer.
12 appsOpen collectionMonty Hall, Simpson, Berkson and base rates: where intuition fails us with numbers.
15 appsOpen collectionPlanets, stars, black holes and space probes: the cosmos to explore.
12 appsOpen collectionLight switches for nerve cells, a telescope made of ice, mirror-image chemistry and interest rates: the 2026 Nobel Prizes to try out.
8 appsOpen collectionPlanted participants, smoke in the room, one dollar for a lie and experimenters who are themselves the subjects: famous social psychology experiments to try yourself.
5 appsOpen collectionRelations, classes, orders, bijections, different infinities and topology: mathematics as the science of structure, hands-on.
11 appsOpen collectionRiemann, P vs NP, Poincaré and other great questions of mathematics, made tangible.
16 appsOpen collectionCounting, arithmetic and shapes: maths for beginners, playful and clear.
6 appsOpen collectionWant to use these apps in your courses? Let’s find out what works for you.
One form, three matters. We answer personally.
We show heyprof on your own material. Tell us what you teach and we will prepare the conversation around it.