Activation Function

/ˌæktɪˈveɪʃən ˈfʌŋkʃən/ac·ti·va·tion func·tionnoun
Neural Networks CoreFoundational

Definition

1.[in machine learning] a mathematical function that introduces non-linearity into the output of a layer, allowing the network to learn complex patterns.

The choice of activation function, such as ReLU, significantly impacts the training stability of the model.

Formal statement

f(x) = max(0, x)

This is the formula for the Rectified Linear Unit (ReLU).

Etymology

From activate, Latin activus 'in action', by way of the biological metaphor of a neuron firing once its input crosses a threshold. The idea dates to the McCulloch-Pitts threshold unit of 1943; the smooth differentiable forms arrived with the backpropagation work of the 1980s.

Synonyms

  • non-linearitysense 1 · Near

See also

References

  • Goodfellow et al. (2016). Deep Learning, ch. 6.MIT Press.