Machine Learning No.5: Neural networks

1. advantage: when number of features is too large, so previous algorithm is not a good way to learn complex nonlinear hypotheses.

2. representation

"activation" of unit i in layer j

matrix of weights controlling function mapping from layer j to layer j+1

3. sample

we have the neural expressions

if network has s_j units in layer j, s_j+1 units in layer j+1, then θ^(j) will be of dimension s_j+1 * (s_j + 1).

4. forward propagation:

add 

5. cost function

L: total no. of layers in network

s_l: no. of units(not counting bias unit) in layer l

6. gradient computation

need code to compute:

backpropagation algorithm:

sample network:

Pace:

7. gradient checking

 

8. random initialization

9. sum.