L2 regularization prevents the accuracy change before and after fitting, and the weight initialization

Recently in the look at Depth model processing NLP text classification. Generally in the writing model, the L2 regularization coefficient is set to 0, not to run regularization. There is also a small trick, is the initialization of some weights, such as the weight of each layer of CNN, as well as the weight of the full connection layer and so on.
In general, these weights may be randomly initialized and conform to a normal distribution. Although the results have little impact, it will certainly affect the convergence and accuracy of the model.
First, two pictures.






The first picture is that I have no home plus the accuracy diagram that prevents overfitting and randomly initializes the weights matrix. The second chapter is that I added the L2 regularization and some method to initialize the weight after the results, found that the model is relatively stable, and there has not been a fitting phenomenon, that is, as the first picture, to a certain stage of the accuracy of the rapid decline.

How home plus will not say, how to initialize the weight, you can see this link. Https://zhuanlan.zhihu.com/p/21560667?refer=intelligentunit