Model Training Shuffle
What is the Shuffle?
shuffle (Chinese meaning: shuffle, confusion). shuffle meaning representation in machine learning and deep learning is the training model data sets disrupting operation.
Original data, may be arranged in a certain order in the sample equalization, the first half of the data as a category, another category of the second half of the data. But after disrupted arrangement of the data it will have a certain randomness in the samples when sequentially read the same time as the possibility to obtain any type of data.
Shuffle model training process needs it?
Shuffle is a training technique, machine learning because of its assumptions and data requirements is to meet the independent and identically distributed . So any sample are required to meet the "randomness." Therefore, under the strong data ** "human" characteristics ** order situation, Shuffle is crucial.
However, the sequence model the model itself, the order of the characteristic data set as the main feature of data, and the time needed to learn this order model laws, can not use Shuffle. Otherwise it will damage the characteristics of the data set.
Shuffle Why is it important?
- Shuffle can prevent jitter model training process is conducive to robust models
assume that the training data is divided into two categories, when Shuffle training has not been, first of all fit the parameters of the model will go to the first class data, when a large amount of continuous data when the (first class) entered the training, the model will result in a first class data over-fitting. After the end of the first class began learning model data for a large number of second class learning data, it would try to approach the second type model data, resulting in a new over-fitting phenomenon. Such repeated training model will fit between the two had lingered, causing jitter model, is not conducive to rapid convergence and convergence training model - Shuffle can prevent over-fitting, learn and make the model more accurate characteristic
learning ability NN network of strong, if the data has not been disrupted, according to the model repeatedly characterized the order of learning data, will soon be reached fitting state and it is possible to learn just the order of feature data. Lack of generalization ability of the model.
Such as: Data 100 in the front 50 of the remaining 50 Class A of Classes B, the model in a very short learning process to learn a boundary point 50, and the first half of the second half of A to B. They did not learn the true class feature.
3 is such that the training set, the validation set, the test set data distribution similar
Question : different types of data is done with shuffle, then partitioned data set; or separate categories corresponding to each operation done?
There is to know little friends can leave a message below
summary
In fact, Shuffle role boils down to two points, on sensitive data sets for randomness
- Enhance the quality of the model
- Enhance the forecasting performance