Input features
Move the sliders or type exact values.
cm
4.37.9
cm
2.04.4
cm
1.06.9
cm
0.12.5
QUICK EXAMPLES
The prediction is computed in your browser using the trained model's actual weights.
FULLY LOCAL INFERENCE / NO API REQUIRED
Inside the mind of a
neural network.
A real, trained AI model. Four measurements in. Three flower species out. Change the inputs, inspect the learned weights, and watch the mathematics come alive.
27Artificial neurons
191Learned parameters
2Hidden layers
93.3%Held-out test accuracy · 28/30
01 / EXPERIMENT
Give the network something to think about.
Measurements in centimeters. The four features come from the Iris flower dataset.
Model prediction
Softmax probability distribution
✿
PREDICTED SPECIES
Iris setosa
The network chose its highest scoring output.
—MODEL SCORE
CLASS PROBABILITIES0—100%
Setosa —
Versicolor —
Virginica —
Model ready for inferenceCPU · JavaScript
02 / SIGNAL STUDIO
One neuron. Every calculation. Your control.
Move a weight, change the bias, or try a different activation. The signal, mathematics, and output all update live.
NEURON ENGINE / LIVE SANDBOX
TRAINED WEIGHTS · REAL MODELFour weighted signals / real neural-network parameters
Engineer the signal
Edit the neuron, not just a decorative diagram.Live Iris measurements ↑
LEARNED WEIGHTS / w₁—w₄drag to edit
WEIGHTED SUM / z—
ACTIVATION / a—
UNDER THE GLASS / ACTUAL FORWARD-PASS MATH
z = Σ(xᵢ × wᵢ) + b = Calculating…
ReLU(z) = max(0, z)
Important: Changing weights or activation here is an educational sandbox. It does not secretly rewrite the trained Iris classifier above. Restore weights to see the real first-layer neuron again.
03 / NETWORK VISUALIZATION
Every connection has a weight.
27 neurons arranged in four layers. Select a neuron in the first hidden layer to inspect its trained calculation, or experiment with its weights in Signal Studio.
INPUT LAYER 04HIDDEN LAYER 01 12HIDDEN LAYER 02 08OUTPUT LAYER 03
04 / UNDER THE HOOD
Not magic. Actual mathematics.
Here is how the selected artificial neuron transforms four normalized inputs into an activation.
Neuron inspector
Click a first-layer neuron above to change this.
z = Σ(xᵢ · wᵢ) + b
WEIGHTED SUM + BIAS—
RELU ACTIVATION—
Inputs are standardized with training-set statistics before reaching the first layer. ReLU(z) = max(0, z).
What's really happening?
From data to decision
01
Normalize
Measurements are standardized with stored training means and standard deviations.
02
Multiply & add
Each layer computes weighted inputs and biases with matrix operations.
03
Activate
Two hidden layers apply ReLU to transform their outputs.
04
Predict
Three output logits are converted to class probabilities using softmax.
Real PyTorch weightsExtracted directly from a model trained from scratch on the classic Iris dataset. No network request, backend, or simulated answers.