Is ReLU a non-linear activation function?

Question

According to this blog post

The purpose of an activation function is to add some kind of non-linear property to the function

The sigmoid is typically used as an activation function of a unit of a neural network in order to introduce non-linearity.

Is ReLU a non-linear activation function? And why? If not, then why is it used as an activation function of neural networks?

Answer 1

ReLU is non-linear by definition

In calculus and related areas, a linear function is a function whose graph is a straight line, that is a polynomial function of degree one or zero.

Since the graph of the ReLU function $f(x) = \max(0,x)$ is not a straight line (equivalently, it cannot be expressed in the form $f(x) = mx + c$), by definition it is not linear.

ReLU is piecewise linear

ReLU is piecewise linear on the bounds $(-\inf,0]$ and $[0,\inf)$:

$$ f(x) = \max(0,x) = \begin{cases} 0 & x \le 0\\ x & x \gt 0\\ \end{cases} $$

But this is still non-linear on the entire domain:

Answer 2

Short Answer: Yes

Visually:

if you see the image from wikipedia, it shown that ReLU (the blue line) is non-Linear (the line is not straight, it turns in 0). You can also check "visual" definition of linear function in wikipedia:

"In calculus and related areas, a linear function is a function whose graph is a straight line"

Mathematically:

Linear function of one variable can be defined as:

$ f(x) = ax + b $

If you plot that function in 2D, it will give you a straight line. Then, the form of linear function with multi variables:

$ f(x_1, x_2, ..., x_n) = a_1x_1 + a_2x_2 + ... + a_nx_n + b $

If you again plot that function in the correct dimension it also give you a straight line. And if you that function carefully, it similar with calculation that happen in a neuron. That's why neuron addition and multiplication is a linear function:

$ f(x_1, x_2, ..., x_n) = w_1x_1 + w_2x_2 + ... + w_nx_n + b $

Adding more layer of linear functions doesn't make the function become "complex" for example, if you have $f(x)$ like below and then you put another layer of linear function $g(x)$ on top of it:

$f(x) = ax + b$

$g(x) = cf(x) + d = cax + cb + d$

as the neural network is trained to find the value of $a,b,c,d$, we can group the constant from the formula above, and then rewrite to:

$h(x) = mx + n$

with $m=ca$ and $n=cb+d$. So without non-linear function the layer of neural network is useless, it only give you another "simple" linear function

ReLU formula is a $f(x)=max(0,x)$, it produces non-linearity as you can't write to linear function format. Using this function will give you "complexity" when you add more layer on top of it.