Network Configuration

Network Configuration

The network configuration controls define the structure and behavior of the neural network.

Architecture

The architecture field uses comma-separated layer sizes.

Example:

2,2,1

This means:

• 2 input neurons.
• 2 hidden neurons.
• 1 output neuron.

Another example:

2,3,2,1

This means:

• 2 input neurons.
• 3 neurons in hidden layer 1.
• 2 neurons in hidden layer 2.
• 1 output neuron.

Activation Function

The activation function is applied to hidden and output layers.

Available options:

• sigmoid
• relu

Loss Function

The current loss function is:

• mse

MSE stands for Mean Squared Error.

Initializer

The initializer defines how weights are initialized.

Available options:

• he
• xavier
• custom

He

He initialization is useful for networks that use ReLU activations.

Xavier

Xavier initialization is commonly used with sigmoid-like activations.

Custom

Custom initialization starts weights and biases at 0 and enables manual editing before training starts.

Manual Weight and Bias Editing

Manual editing is only available when the initializer is set to custom.

In this mode, the sidebar shows:

• Weight selector.
• Weight value input.
• Weight reset button.
• Bias selector.
• Bias value input.
• Bias reset button.

After training starts, manual editing is disabled until the network is reset.