Backpropagation Puzzle Game

Network laboratory

Configure the network, task, training method, and puzzle rules.

Dataset scenario

Difficulty

Learning mode

Puzzle type

Input neurons

2 inputs

Hidden layers

1 hidden layer

Hidden layer 1 neurons

3 neurons

Hidden layer 2 neurons

2 neurons

Output neurons

1 output

Hidden activation

Output activation

Loss function

Learning rate

η = 0.120

Batch size

Gradient clipping

Weight decay

Loss—

Prediction—

Target—

Gradient norm—

Score0

Epoch0

Interactive neural network canvas

Click neurons and connections to inspect forward values, deltas, gradients, weights, and biases.

positive negative selected frozen

Network ready

Manual parameter workshop

Edit weights and biases, preview the new loss, then apply or restore changes.

Parameter	Current	Gradient	Proposed	State

Plotly diagnostics

Track loss, predictions, parameter movement, gradient flow, and activation behaviour.

Puzzle station

Solve the current backpropagation challenge.

Untimed

Gradient tracing puzzle

Run a forward and backward pass to generate a puzzle.

Complete a network step, then answer the generated question.

Puzzle progress0 / 10

Calculation inspector

Selected network element and exact chain-rule calculation.

Select a node or connection on the canvas.

Forward: z = Σ(w × a) + b Activation: a = f(z) Backward: δ = ∂L/∂a × f′(z) Weight gradient: ∂L/∂w = δ × a(previous)

Training sample

Change examples or compare sample gradients.

Progress and achievements

Saved automatically in this browser.

Accuracy0%

Hints used0

Best score0

Accessibility and display

Adjust motion, contrast, precision, labels, and sound.

High contrast

Reduced animation

Sound feedback

Scientific notation

Decimal precision

Learning guide

Formulas, keyboard controls, and common gradient problems.

Backpropagation steps

Run the forward pass.
Calculate the output loss.
Find output deltas.
Propagate deltas backward.
Calculate weight and bias gradients.
Update parameters and compare loss.

Keyboard shortcuts

F forward · B backward · U update · Space train or pause · N new puzzle

Gradient problem clues

Tiny deep-layer gradients suggest vanishing. Extremely large norms suggest exploding. ReLU units with zero activation and zero derivative may be dead.

Experiment comparisons

Compare learning rates, activation functions, clipping choices, and manual versus automatic updates.

Generated 2026-08-11. All datasets and predictions are fictional and educational.

Network laboratory

Interactive neural network canvas

Manual parameter workshop

Plotly diagnostics

Experiment comparisons

Related Calculators