stem/AI/Neural Networks/Activation Functions.md

tags

• Limits output values
• Squashing function

Threshold

• For binary functions
• Not differentiable
  • Sharp rise
• Heaviside function
• Unipolar
  • 0 <-> +1
• Bipolar
  • -1 <-> +1

Sigmoid

• Logistic function
• Normalises
• Introduces non-linearity
• Alternative is tanh
  • -1 <-> +1
• Easy to take derivative

\frac d {dx} \sigma(x)=
\frac d {dx} \left[ 
\frac 1 {1+e^{-x}}
\right]
=\sigma(x)\cdot(1-\sigma(x))$$

![sigmoid](../../img/sigmoid.png)

### Derivative

$$y_j(n)=\varphi_j(v_j(n))=
\frac 1 {1+e^{-v_j(n)}}$$
$$\frac{\partial y_j(n)}{\partial v_j(n)}=
\varphi_j'(v_j(n))=
\frac{e^{-v_j(n)}}{(1+e^{-v_j(n)})^2}=
y_j(n)(1-y_j(n))$$
- Nice derivative
- Max value of $\varphi_j'(v_j(n))$ occurs when $y_j(n)=0.5$
- Min value of 0 when $y_j=0$ or $1$
- Initial [weights](Weight%20Init.md) chosen so not saturated at 0 or 1

If $y=\frac u v$
Where $u$ and $v$ are differential functions

$$\frac{dy}{dx}=\frac d {dx}\left(\frac u v\right)$$

$$\frac{dy}{dx}=
\frac {v \frac d {dx}(u) - u\frac d {dx}(v)} {v^2}$$

# ReLu
Rectilinear
- For deep networks
- $y=max(0,x)$
- CNNs
	- Breaks associativity of successive [convolutions](../../Signal%20Proc/Convolution.md)
		- Critical for learning complex functions
	- Sometimes small scalar for negative
		- Leaky ReLu

![relu](../../img/relu.png)

# SoftMax
- Output is per-class vector of likelihoods
	- Should be normalised into probability vector

## AlexNet
$$f(x_i)=\frac{\text{exp}(x_i)}{\sum_{j=1}^{1000}\text{exp}(x_j)}$$