This document discusses multi-layer perceptrons (MLPs), including their advantages over single-layer perceptrons. MLPs can classify problems that single-layer perceptrons cannot by using multiple hidden layers between the input and output layers. MLPs are trained using an error-based learning method called backpropagation, which calculates errors between the target and actual output values and adjusts weights in the network accordingly starting from the output layer and propagating backwards. MLPs are well-suited for parallel processing architectures.

1. Multi Layer Perceptron
Computer & Robot Vision Lab
Sung -ju Kim
goddoe2@gmail.com

2. Content
• Why Neural Net came back?
• Single Layer Perceptron
• Multi Layer Perceptron
• Traffic Sign Lane Guessing

3. Why Neural Net Came back?
Stallkamp, Johannes, et al. "Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition." Neural networks 32 (2012): 323-332.
Face net: 99.6%
Deep Face : 97.25%

4. Single Layer Perceptron
Feed Forward

5. Activation Functions
step function sign function identity function
sigmoid function hyper tangent function

6. Limitation of Single Layer Perceptron
But Single Layer Perceptron
cannot classify XOR Problem

7. Multi Layer Perceptron

8. Multi Layer Perceptron
Hidden
Layer
Output
Layer
Input
Layer

9. Ex)
✕
✕
Feed Forward

10. How to Learn Perceptron?
Error Function
Error Function
Target Value
Output Value

11. How to Learn Perceptron?
Delta Learning Rule
E
(weight)

12. How to Learn Perceptron?
Delta Learning Rule
new weight
current weight
learning rate
Error Function

13. How to Learn Perceptron?
Delta Learning Rule
E
(weight)
new weight
current weight
learning rate
Error Function

14. Back Propagation
Hidden
Layer
Output
Layer
Input
Layer

15. new weight
current weight
learning rate
Error Function
Target Value
Output Value

16. Back Propagation
Hidden
Layer
Output
Layer
Input
Layer

17. new weight
current weight
learning rate
Error Function

18. Good for Design in Parallels Architecture