Use this tool to inspect the exact scalar or vector transformation applied between neural-network layers, reproduce a hand calculation, or compare how activations treat negative, zero, and large positive inputs.

Values

Raw scores.

The implemented rule is: Convert a list of scores into a probability distribution with the numerically stable softmax.^[1]

Probabilities

Probabilities whose mathematical sum is one.

An activation value alone does not predict training quality. Gradient behavior, initialization, normalization, architecture, numeric precision, loss function, and the distribution of inputs all affect whether an activation is suitable. Very large magnitudes can also expose floating-point limits even when a stable formula is used.

Plot the activation and its derivative across the expected input range, then compare validation behavior with another activation under the same initialization and training settings. Use the softmax tool only when the outputs must be normalized jointly rather than transformed independently.

• Leaky ReLU

  Apply the leaky ReLU activation, scaling negatives by a small slope.

• ReLU

  Apply the rectified linear activation, flooring negatives at zero.

• Sigmoid

  Apply the logistic sigmoid activation.