Activation function

Logistic activation function

In artificial neural networks, the activation function of a node defines the output of that node given an input or set of inputs. A standard integrated circuit can be seen as a digital network of activation functions that can be "ON" (1) or "OFF" (0), depending on input. This is similar to the linear perceptron in neural networks. However, only nonlinear activation functions allow such networks to compute nontrivial problems using only a small number of nodes, and such activation functions are called nonlinearities.^[1]

Classification of activation functions[]

The most common activation functions can be divided in three categories: ridge functions, radial functions and fold functions.

Ridge activation functions[]

Ridge functions are multivariate functions acting on a linear combination of the input variables. Often used examples include:

• Linear activation: ${\displaystyle \phi (\mathbf {v} )=a+\mathbf {v} '\mathbf {b} }$,
• ReLU activation: ${\displaystyle \phi (\mathbf {v} )=\max(0,a+\mathbf {v} '\mathbf {b} )}$,
• Heaviside activation: ${\displaystyle \phi (\mathbf {v} )=1_{a+\mathbf {v} '\mathbf {b} >0}}$,
• Logistic activation: ${\displaystyle \phi (\mathbf {v} )=(1+\exp(-a-\mathbf {v} '\mathbf {b} ))^{-1}}$.

In biologically inspired neural networks, the activation function is usually an abstraction representing the rate of action potential firing in the cell.^[2] In its simplest form, this function is binary—that is, either the neuron is firing or not. The function looks like ${\displaystyle \phi (\mathbf {v} )=U(a+\mathbf {v} '\mathbf {b} )}$, where ${\displaystyle U}$ is the Heaviside step function.

A line of positive slope may be used to reflect the increase in firing rate that occurs as input current increases. Such a function would be of the form ${\displaystyle \phi (\mathbf {v} )=a+\mathbf {v} '\mathbf {b} }$.

Rectified linear unit and Gaussian error linear unit activation functions

Neurons also cannot fire faster than a certain rate, motivating sigmoid activation functions whose range is a finite interval.

Radial activation functions[]

A special class of activation functions known as radial basis functions (RBFs) are used in RBF networks, which are extremely efficient as universal function approximators. These activation functions can take many forms such as:

• Gaussian: ${\displaystyle \,\phi (\mathbf {v} )=\exp \left(-{\frac {\|\mathbf {v} -\mathbf {c} \|^{2}}{2\sigma ^{2}}}\right)}$
• Multiquadratics: ${\displaystyle \,\phi (\mathbf {v} )={\sqrt {\|\mathbf {v} -\mathbf {c} \|^{2}+a^{2}}}}$

where ${\displaystyle \mathbf {c} }$ is the vector representing the function center and ${\displaystyle a}$ and ${\displaystyle \sigma }$ are parameters affecting the spread of the radius.

Folding activation functions[]

Folding activation functions are extensively used in the pooling layers in convolutional neural networks, and in output layers of multiclass classification networks. These activations perform aggregation over the inputs, such as taking the mean, minimum or maximum. In multiclass classification the softmax activation is often used.

Comparison of activation functions[]

There are numerous activation functions. Hinton et al.'s seminal 2012 paper on automatic speech recognition uses a logistic sigmoid activation function.^[3] The seminal 2012 AlexNet computer vision architecture uses the ReLU activation function, as did the seminal 2015 computer vision architecture ResNet. The seminal 2018 language processing model BERT uses a smooth version of the ReLU, the GELU.^[4]

Aside from their empirical performance, activation functions also have different mathematical properties:

Nonlinear

When the activation function is non-linear, then a two-layer neural network can be proven to be a universal function approximator.^[5] This is known as the Universal Approximation Theorem. The identity activation function does not satisfy this property. When multiple layers use the identity activation function, the entire network is equivalent to a single-layer model.

Range

When the range of the activation function is finite, gradient-based training methods tend to be more stable, because pattern presentations significantly affect only limited weights. When the range is infinite, training is generally more efficient because pattern presentations significantly affect most of the weights. In the latter case, smaller learning rates are typically necessary.^{[citation needed]}

Continuously differentiable

This property is desirable (ReLU is not continuously differentiable and has some issues with gradient-based optimization, but it is still possible) for enabling gradient-based optimization methods. The binary step activation function is not differentiable at 0, and it differentiates to 0 for all other values, so gradient-based methods can make no progress with it.^[6]

Sign equivalence to identity function[]

Two real valued functions f and g are said to be sign equivalent if ${\displaystyle sign(f(z))=sign(g(z))}$^[7] for all values of z in the domain. Where sign is the signum function. Activation functions like tanh, Leaky ReLU, GELU, ELU, Swish and Mish are sign equivalent to the identity function and cannot learn the XOR function with a single neuron.^[8] The output of a single neuron or its activation is ${\displaystyle a=g(z)=g({\boldsymbol {w}}^{T}{\boldsymbol {x}}+b)}$, where g is the activation function. The decision boundary for a single neuron is the set of points that elicit an output of zero. Thus the decision boundary for a neuron using any of the activation functions sign equivalent to the identity function is a single hyperplane. However oscillatory activation functions can have many zeros and hence a single neuron can have multiple hyperplanes as part of its decision boundary. Although multilayer networks are needed to achieve nonlinear decision boundaries, the use of oscillatory activation functions allows even single neurons to exhibit nonlinear decision boundaries.^[7]

These properties do not decisively influence performance, nor are they the only mathematical properties that may be useful. For instance, the strictly positive range of the softplus makes it suitable for predicting variances in variational autoencoders.

The following table compares the properties of several activation functions that are functions of one fold x from the previous layer or layers:

Name	Plot	Function, ${\displaystyle f(x)}$	Derivative of ${\displaystyle f}$, ${\displaystyle f'(x)}$	Range	Order of continuity
Identity		${\displaystyle x}$	${\displaystyle 1}$	${\displaystyle (-\infty ,\infty )}$	${\displaystyle C^{\infty }}$
Binary step		${\displaystyle {\begin{cases}0&{\text{if }}x<0\\1&{\text{if }}x\geq 0\end{cases}}}$	${\displaystyle {\begin{cases}0&{\text{if }}x\neq 0\\{\text{undefined}}&{\text{if }}x=0\end{cases}}}$	${\displaystyle \{0,1\}}$	${\displaystyle C^{-1}}$
Logistic, sigmoid, or soft step		${\displaystyle \sigma (x)={\frac {1}{1+e^{-x}}}}$[1]	${\displaystyle f(x)(1-f(x))}$	${\displaystyle (0,1)}$	${\displaystyle C^{\infty }}$
Hyperbolic tangent (tanh)		${\displaystyle \tanh(x)={\frac {e^{x}-e^{-x}}{e^{x}+e^{-x}}}}$	${\displaystyle 1-f(x)^{2}}$	${\displaystyle (-1,1)}$	${\displaystyle C^{\infty }}$
Rectified linear unit (ReLU)[9]		${\displaystyle {\begin{aligned}&{\begin{cases}0&{\text{if }}x\leq 0\\x&{\text{if }}x>0\end{cases}}\\{}={}&\max\{0,x\}=x{\textbf {1}}_{x>0}\end{aligned}}}$	${\displaystyle {\begin{cases}0&{\text{if }}x<0\\1&{\text{if }}x>0\\{\text{undefined}}&{\text{if }}x=0\end{cases}}}$	${\displaystyle [0,\infty )}$	${\displaystyle C^{0}}$
Gaussian Error Linear Unit (GELU)[4]		${\displaystyle {\begin{aligned}&{\frac {1}{2}}x\left(1+{\text{erf}}\left({\frac {x}{\sqrt {2}}}\right)\right)\\{}={}&x\Phi (x)\end{aligned}}}$	${\displaystyle \Phi (x)+x\phi (x)}$	${\displaystyle (-0.17\ldots ,\infty )}$	${\displaystyle C^{\infty }}$
Softplus[10]		${\displaystyle \ln \left(1+e^{x}\right)}$	${\displaystyle {\frac {1}{1+e^{-x}}}}$	${\displaystyle (0,\infty )}$	${\displaystyle C^{\infty }}$
Exponential linear unit (ELU)[11]		${\displaystyle {\begin{cases}\alpha \left(e^{x}-1\right)&{\text{if }}x\leq 0\\x&{\text{if }}x>0\end{cases}}}$ with parameter ${\displaystyle \alpha }$	${\displaystyle {\begin{cases}\alpha e^{x}&{\text{if }}x<0\\1&{\text{if }}x>0\\1&{\text{if }}x=0{\text{ and }}\alpha =1\end{cases}}}$	${\displaystyle (-\alpha ,\infty )}$	${\displaystyle {\begin{cases}C^{1}&{\text{if }}\alpha =1\\C^{0}&{\text{otherwise}}\end{cases}}}$
Scaled exponential linear unit (SELU)[12]		${\displaystyle \lambda {\begin{cases}\alpha (e^{x}-1)&{\text{if }}x<0\\x&{\text{if }}x\geq 0\end{cases}}}$ with parameters ${\displaystyle \lambda =1.0507}$ and ${\displaystyle \alpha =1.67326}$	${\displaystyle \lambda {\begin{cases}\alpha e^{x}&{\text{if }}x<0\\1&{\text{if }}x\geq 0\end{cases}}}$	${\displaystyle (-\lambda \alpha ,\infty )}$	${\displaystyle C^{0}}$
Leaky rectified linear unit (Leaky ReLU)[13]		${\displaystyle {\begin{cases}0.01x&{\text{if }}x<0\\x&{\text{if }}x\geq 0\end{cases}}}$	${\displaystyle {\begin{cases}0.01&{\text{if }}x<0\\1&{\text{if }}x\geq 0\end{cases}}}$	${\displaystyle (-\infty ,\infty )}$	${\displaystyle C^{0}}$
Parameteric rectified linear unit (PReLU)[14]		${\displaystyle {\begin{cases}\alpha x&{\text{if }}x<0\\x&{\text{if }}x\geq 0\end{cases}}}$ with parameter ${\displaystyle \alpha }$	${\displaystyle {\begin{cases}\alpha &{\text{if }}x<0\\1&{\text{if }}x\geq 0\end{cases}}}$	${\displaystyle (-\infty ,\infty )}$[2]	${\displaystyle C^{0}}$
Sigmoid linear unit (SiLU,[4] Sigmoid shrinkage,[15] SiL,[16] or Swish-‍1[17])		${\displaystyle {\frac {x}{1+e^{-x}}}}$	${\displaystyle {\frac {1+e^{-x}+xe^{-x}}{\left(1+e^{-x}\right)^{2}}}}$	${\displaystyle [-0.278\ldots ,\infty )}$	${\displaystyle C^{\infty }}$
Mish [18]		${\displaystyle x\tanh(\ln(1+e^{x}))}$	${\displaystyle {\frac {(e^{x}(4e^{2x}+e^{3x}+4(1+x)+e^{x}(6+4x)))}{(2+2e^{x}+e^{2x})^{2}}}}$	${\displaystyle [-0.308\ldots ,\infty )}$	${\displaystyle C^{\infty }}$
Gaussian		${\displaystyle e^{-x^{2}}}$	${\displaystyle -2xe^{-x^{2}}}$	${\displaystyle (0,1]}$	${\displaystyle C^{\infty }}$
Growing Cosine Unit (GCU)[7]		${\displaystyle x\cos(x)}$	${\displaystyle \cos(x)-x\sin(x)}$	${\displaystyle (-\infty ,\infty )}$	${\displaystyle C^{\infty }}$

^ Here, ${\displaystyle \sigma }$ is the logistic function.

^ ${\displaystyle \alpha >0}$ for the range to hold true.

The following table lists activation functions that are not functions of a single fold x from the previous layer or layers:

Name	Equation, ${\displaystyle f_{i}\left({\vec {x}}\right)}$	Derivatives, ${\displaystyle {\frac {\partial f_{i}\left({\vec {x}}\right)}{\partial x_{j}}}}$	Range	Order of continuity
Softmax	${\displaystyle {\frac {e^{x_{i}}}{\sum _{j=1}^{J}e^{x_{j}}}}}$    for i = 1, …, J	${\displaystyle f_{i}\left({\vec {x}}\right)\left(\delta _{ij}-f_{j}\left({\vec {x}}\right)\right)}$[3][4]	${\displaystyle (0,1)}$	${\displaystyle C^{\infty }}$
Maxout[19]	${\displaystyle \max _{i}x_{i}}$	${\displaystyle {\begin{cases}1&{\text{if }}j={\underset {i}{\operatorname {argmax} }}\,x_{i}\\0&{\text{if }}j\neq {\underset {i}{\operatorname {argmax} }}\,x_{i}\end{cases}}}$	${\displaystyle (-\infty ,\infty )}$	${\displaystyle C^{0}}$

^ Here, ${\displaystyle \delta _{ij}}$ is the Kronecker delta.

^ For instance, ${\displaystyle j}$ could be iterating through the number of kernels of the previous neural network layer while ${\displaystyle i}$ iterates through the number of kernels of the current layer.

See also[]

• Logistic function
• Rectifier (neural networks)
• Stability (learning theory)
• Softmax function

References[]