Formula used
Weighted input: z[l] = W[l]a[l−1] + b[l]
Activation: a[l] = f[l](z[l])
Softmax: pi = exp(zi/T) ÷ Σ exp(zj/T)
How to use
- Set the input feature count and total layers.
- Enter one input sample on each line.
- Configure neurons, activations, weights, and biases.
- Select the prediction mode and optional scaling.
- Submit the form to inspect every network step.
Example data
| Element | Example | Meaning |
|---|---|---|
| Input | [0.5, −1.0] | Two input features. |
| Hidden weights | [[0.5, −0.3], [0.8, 0.2]] | Two hidden neurons. |
| Hidden bias | [0.1, −0.2] | One bias per neuron. |
| Activation | ReLU | Negative values become zero. |
| Output | Softmax probabilities | Class probabilities sum to one. |
Forward propagation learning notes
Forward propagation moves values from inputs toward outputs. Each layer transforms the previous layer. Weights control each connection's influence.
Biases shift neuron values before activation. Activation functions introduce useful nonlinearity. Softmax converts logits into class probabilities.
Inference normally disables dropout. Training simulation can apply random dropout masks. Batch normalisation uses stored inference statistics here.
Residual connections require matching vector sizes. They can improve signal flow. Invalid dimensions are reported before computation.
Frequently asked questions
What is forward propagation?
It calculates neural network outputs from inputs, weights, biases, and activations.
Why must matrix dimensions match?
Each neuron needs one weight for every value entering its layer.
When should softmax be used?
Use it for mutually exclusive multiclass classification outputs.
Can this calculator process batches?
Yes. Enter multiple input samples on separate lines.
What does temperature scaling change?
Lower temperatures sharpen probabilities. Higher temperatures soften them.
Is dropout used during inference?
No. Dropout is disabled unless training simulation is selected.
What custom expressions are allowed?
Use x, numbers, arithmetic, parentheses, and supported mathematical functions.
How are residual connections selected?
Choose an earlier activation index with the same output length.
Can results be exported?
Yes. Export CSV, JSON, PDF, printouts, and network configurations.
Does this calculator train a model?
No. It demonstrates forward propagation without backpropagation or optimisation.