Linear

The Linear rule computes a neuron’s activation by scaling its input with a constant slope, optionally adding noise and applying a clipping method. This rule is commonly used in simple neural models and serves as a base for other activation types.

The basic update rule is:

\[a = \text{slope} \cdot x\]

Where:

  • \(a\) is the neuron’s activation,
  • \(x\) is the total input,
  • The slope scales the response.

Different clipping types determine how this result is transformed:

  • No Clipping: \(a = \text{slope} \cdot x\)
  • Relu: \(a = \max(0, \text{slope} \cdot x)\)
  • Piecewise Linear: \(a = \min(\text{upperBound}, \max(\text{lowerBound}, \text{slope} \cdot x))\)

Noise can optionally be added to the activation before clipping.

Parameters

  • Slope: The multiplier applied to the neuron’s input to compute its activation.
  • Type: The clipping strategy applied to the activation. Options are:
    • No Clipping
    • Piecewise Linear
    • ReLU

For all other parameters, see common neuron properties