TRAINING LOSS
—
Binary cross-entropyAwaiting run
INTERACTIVE MACHINE LEARNING
See how a neural network learns.
Build a tiny neural net. Train it on real patterns. Inspect its weights, activations and gradients—one experiment at a time.
TRAIN ACCURACY
—%
On current datasetBaseline
EPOCHS COMPLETED
0
Training progressTarget: 120
TRAINABLE PARAMETERS
105
Weights + biases2–8–8–1
01 / LEARNED REPRESENTATION
Decision boundary LIVE
See the model's prediction across the input space.
Class 0 Class 1
FEATURE 2
FEATURE 1
◉ Each dot is one labeled training examplePredicted probability 0 → 1
02 / TRAINING CONTROL
Configure & train
Change a setting and observe the effect.
A classic non-linear decision problem.
80600 examples
CleanNoisy
A fixed seed makes experiments easier to reproduce.
2→→1
2 inputs · 1 output · 1–4 hidden layers, 2–32 neurons each
ConservativeAggressive
20300 epochs
Ready to learn0%
03 / INSIDE THE MODEL
Network architecture
Explore how information moves from input to prediction.
INPUT LAYERHIDDEN LAYERSOUTPUT
Initialising the network…
Positive NegativeLine width = magnitudeWeight values
04 / WEIGHT INSPECTOR
Neuron inspector
Select a node to inspect its parameters.
◎Select a neuron
Click any node in the network map to inspect its activation, bias and incoming weights.
05 / OPTIMISATION
Learning curve
Binary cross-entropy across training epochs.
Train the model to build the learning curve.
CONCEPT CARD
What is a decision boundary?
A decision boundary is the line or surface a classifier uses to separate classes. A neural network creates it by composing simple mathematical transformations. Watch this map evolve as the loss decreases.
ŷ = σ(W₂ · φ(W₁x + b₁) + b₂)
CONTROLLED EXPERIMENTS
Experiment lab.
Compare learning choices on the same dataset. Keep the architecture and data fixed, then change one variable at a time.
CURRENT EXPERIMENT
XOR classification
The active dataset will be shared across each comparison run.
Make a fair comparison
Each configuration starts from its own initial weights and trains on the same dataset. Results are educational estimates, not a scientific benchmark suite.
Configuration matrix
Four complementary baselines, one consistent dataset.
01BASELINE
Adam + Tanh
Adaptive step sizes with smooth hidden activations.
Results appear here after a run.
02CLASSIC
SGD + Tanh
A simple gradient step with no adaptive moments.
Results appear here after a run.
03SPARSE
Adam + ReLU
Adaptive optimisation with a piecewise-linear activation.
Results appear here after a run.
04INERTIA
Momentum + Tanh
Gradient descent with a moving average of updates.
Results appear here after a run.
RESULTS
Outcome comparison
Lower loss and higher accuracy are better, but also inspect stability and convergence.
Run a comparison to populate the chart.
| Configuration | Final loss | Accuracy | Parameters | Time |
|---|---|---|---|---|
| No experiment results yet. | ||||
INTUITION BEFORE EQUATIONS
Learn the building blocks.
A hands-on field guide to the pieces that make a small neural network work. Change the model and watch these ideas become visible.
∑
THE CORE IDEA
Learn a function from examples.
A neural network is a chain of parameterised transformations. Training adjusts those parameters so predictions become more useful on examples with known labels.
01
Neurons & layers
Each neuron combines inputs with weights, adds a bias, and applies an activation function. Layers compose these small operations into a flexible model.
z = w · x + b
Try thisAdd a hidden layer or change the neuron count in Playground.
02
Activation functions
Non-linear activations allow networks to represent curved and complex boundaries. Without non-linearity, stacked dense layers collapse into one linear transformation.
a = φ(z)
Try thisCompare Tanh, ReLU, Leaky ReLU and Sigmoid.
03
Loss & gradients
Loss measures the cost of predictions. Backpropagation applies the chain rule to estimate how each parameter contributes to that cost.
∂L / ∂w
Try thisSwitch the network map to Gradients after training.
04
Optimisation
An optimiser turns gradients into parameter updates. SGD, Momentum and Adam use different rules, each with trade-offs in speed and stability.
w ← w − ηg
Try thisRun the same dataset with different optimisers.
05
Generalisation
A model can memorise examples without learning a reliable pattern. Noise, dataset shape and model capacity all influence how a boundary behaves.
Train ≠ unseen data
Try thisIncrease dataset noise or select the two-spiral dataset.
06
Experimental thinking
Change one variable, record the outcome, and repeat from comparable initial conditions. A single run is an illustration, not proof of a general claim.
Hypothesis → test → evidence
Try thisUse Experiment Lab to compare four training configurations.
QUICK REFERENCE
Glossary
Small words, big ideas.
Epoch
One complete pass over all training examples.
Batch
A subset of examples used to calculate one update.
Learning rate
The step scale used when updating model parameters.
Bias
A trainable offset that shifts a neuron's response.
Parameter
A learnable value, such as a weight or bias.
Decision boundary
The region where the predicted class changes.
MODEL REFERENCE
Model guide.
A transparent map of the current implementation, its math, its limits, and ways to reproduce an experiment.
MODEL CONTRACT
What this lab trains
A fully connected multilayer perceptron for two-dimensional, binary classification. The input layer has two features and the output is a single sigmoid probability.
input (2) → hidden layers → sigmoid (1)
Hidden widths are configurable (1–4 layers, 2–32 neurons per layer). The demo uses synthetic data and runs training locally in the browser.
OBJECTIVE
Binary cross-entropy
The output probability is compared with a binary target. Numerical clamping keeps the logarithm finite near 0 and 1.
L = −[y log(p) + (1−y) log(1−p)]
With sigmoid output and binary cross-entropy, the output-logit derivative simplifies to p − y.
OPTIMISERS
Three update rules
SGD applies the current gradient. Momentum smooths updates with a running velocity. Adam adapts each parameter's update using first and second gradient moments.
SGD: w ← w − ηg
All optimisers use the same binary objective and training examples for a fairer, though still illustrative, comparison.
REPRODUCIBILITY
Export and repeat
Export the current model weights as JSON, or the exact synthetic points as CSV. The browser model file includes its architecture, activation, optimizer, learning rate and parameter arrays.
PYTHON TRACK
NumPy + Matplotlib
The repository also contains a small NumPy implementation and a command-line experiment script. Use it for saved metrics, reproducible seeds and decision-boundary plots.
python -m pip install -r requirements.txt
python scripts/train_numpy.py --dataset xor
python scripts/train_numpy.py --dataset xor
Browser experiments and Python runs are separate execution paths; their results are not expected to match exactly because their random-number streams differ.
LIMITS & SAFETY
Learning demo, not a benchmark
Small synthetic datasets are intentionally simple. Training accuracy is measured on the generated training examples and should not be interpreted as generalisation performance.
For research claims, use separate validation/test data, multiple seeds, explicit baselines, and report uncertainty.