An Essay on Overfit

See @ Love Coco-love life forwarded articles. Look at it a little bit and record it here.

Overfit is an important concept of machine learning. In the narrow sense can be defined as the model is too complex, resulting in the model generalization is not good enough. I think that a broad definition should be used as a general generalization, which can be defined as overfit. From my definition, it can be thought that this article revolves around Overfit, and discusses from three aspects the mistakes that the novice veteran will make.

The first section outlines Overfit, which can be seen as discussing overfit from the complexity of modeling, and the more complex your model is, the easier it is to overfit. Where Overfit is memorizing rather than learning analogy is apt. We need to learn the law through something rather than simply remembering someting. I add that overfit can introduce regulatization through Bayes ' prior to improve generalization.

The second part discusses Overfit from the point of view of data, and the data you use for training will also bring Overfit. This section can be divided into two sections: first of all, bias training data will certainly bias your predictions, generalization is not good enough, overfit with the resulting. Next, leakage training data, will introduce the test data to the model, equal to your white test, cross-validating hyper-param Void, can not generalization, so overfit again. Indeed, many papers did not pay attention to this point, Microsoft has been the joke.

An Essay on Overfit