Popular understanding of deviation and variance in machine learning

One 、 Reference source

What is the difference between deviation and variance ？ - You know

Two 、 Easy to understand

2.1 Basic knowledge of

（1） Deviation is the degree to which a model deviates from real data .

（2） Variance measures the dispersion of different models in a model set .

（3） The real data does not change with the change of some characteristics , That is, the hypothetical result A Mainly characterized by a And characteristics b decision , Study only the characteristics a And characteristics b For the result A Influence , But not the characteristics a=x1, features b=x2, result A It must be y, This is because there are other secondary factors in “ exorcise ”.

（4） be based on （3）, The data in the real world should be f(x)±ε, there ε Is the noise value . generally speaking , Noise produced by secondary factors , It is a normal distribution .

2.2 Graphical deviation and variance

If the real world model is understood as a curve , Then the curve obtained according to the principle of minimum loss function is a straight line in the middle , Noise has little effect on it .

  If the real world model is understood as a conic , Because the model library of conic contains conic , Therefore, it is said that its learning ability becomes stronger , Then you will inevitably learn “ noise ”. Noise is uncertain , Very discrete , Learn a characteristic of instability , Finally, the inevitable variance becomes larger ！