[lihongyi] notes on deep learning of machine learning -- Introduction to training model clearance

Detailed video stamp --> Machine learning task Introduction

If the training results are not satisfactory ： First check training data Of loss , Look at you model stay training data Did you learn it in school , Then go to see testing Result .

model bias

It's yours model Too simple ,model Is not elastic enough .

resolvent ： To reset model

1. Add input features
2. Deep Learning ： more neurons,layers.（ increase model The elasticity of ）

Optimization Issue

namely gradient descent The algorithm cannot be found loss Low function .

Identify whether it is optimization issue Methods

You can tell by comparing different models model Whether the elasticity of is large enough ： see training data Upper loss value .

overfitting

That is, the error rate in the training data is low , But in the case of high error rate in the predicted data .

Use an extreme example to illustrate Overfitting The situation of ：

General examples ：model Is too elastic

solve overfitting Methods ：

1. Add training materials （ The most effective and useful method ）
（1） Add data by gathering more information
（2）data augmentation（ Create new materials according to your own understanding , For example, when doing image recognition , You can reverse the picture from left to right 、 Zoom in and so on , But it must be handled properly ）

2. to model Some restrictions （ For example, it is limited to a conic ）（ How to limit the method to be added later ）

The following figure shows the limitations model Five ways of doing this ：

But don't give too many restrictions , Otherwise it will become model bias It's a question of ！

Dataset cutting

A toolkit for calculating differential ：