The algorithms are trying to mimic brain.
It is very popular in 80s and early 90s. However, popular just diminished in late 90s.
It recently become popular becasue it 's the basement of the artificial intellengence.
So, the pictures shows a neuro, which helps people finish their daily life activities.
dendrite to get informations from other neuro, which can be consider as input.
cell body can deal with the information from other input and convert to other information pass to other neuro.
Axon helps pass the information to other neuro, which can be consider as ouput.
The whole process can be summaried as the second pictures. and it's also the constructions of neural network algorithms
For supervise tasks, if the output is numeric, it can do predictions; if the output is categorical variables, it can do classifications.
For unsupervise learning, if the output node equals to input node, the hidden node information can used to present information at input node. So, it can do the feature extraction as PCA.
Before the neural network process, it need standaradize the data. Because the neural network are data depend algorithms. The scale of variables can impact the performance of the result.
So, before the Neural Network process, scale and standard process need to be apply to the data.
what input and output layer be?
Input layer : feature you want to use in predictions and classifications
Output layer: the target variable
How we calculate the weight and update weight
Using forward algorithms, minimize the loss to get the initial weight and used backward prorogation to update weight
How can we get output from hidden layer?
Activation function help us get result from hidden layer
Loss function:
To evaluate how model performs, like gini and impurity in decision tree.
The nerual network algorithms is about to minimize Loss
Learning rate:
An important parameter in neural network
range from 0 to 1
if learning rate is too big, you may miss the lowest point. if learning rate is too small, it takes time to find the point.
Activation function:
get point from input to output
usually considered as blackbox procedure
common activation functions:
Tanh, Relu, Sigmoid(classification) , Softmax( multi class classfication)
Pros:
Good predictive performance
Have high tolerance to noisy data
Ability to capture highly complicated relationships between the predictors and an outcome variable
Cons
Blackbox reputation
Extrapolation is a serious danger
Does not have a built-in variable selection mechanism
Relies heavily on having sufficient data for training purposes
Computation time
I am still thinking put the example to above to help understanding.
Since it's kind of hard to understand what happens in Neural Network Procedures
But anyways, hope you quite enjoy this part
Happy neural network!