This article is for learning "deep learning Getting Started" chapter in learning about neural networks (written by Japanese scholars Saito Kang Yi Lu Yujie translated by a) to summarize some of the harvest.
Neural network learning book presents four steps: 1.mini-batch 2. calculate the gradient parameter update 4. 3. Repeat the previous step
1. learning neural network from digital handwriting recognition project
The so-called "learning from data" refers to data from the # automatic decision weight #. When solving simple problems, using a simple neural network, the network weights may be artificially set manually input information to extract a particular feature. But in the actual neural network parameters are often thousands, even millions of possible heavy weights, this time artificially set manually is an impossible task. Therefore, when you can get to learn from the data network weights, the computer is a smart thing. This article analyzed the neural network learning through practical projects handwritten digit recognition . Effective use of data, there are two solutions: 1 start the image feature amount is extracted, then the machine learning techniques to learn these feature amounts of patterns. The first step: The method of extracting a feature quantity from the image artificially setting image feature extraction filter, filters have used: SIFT, SURF HOG and the like, the use of these filters the image data into a vector, the final conversion after the machine learning vector SVM, KNN classifier learning and the like. The disadvantage of this solution is that the feature extraction filters need to artificially designed manually, it is necessary to set different filters for different problems. 2. The method of neural networks (deep learning) is entirely data to train weight, find the appropriate characteristics, and make judgments. FIG P83 in FIG. 1 on page 4-2 of the book has a more simple and intuitive compared the difference between the two schemes. Advantage of neural networks is that all the problems can be solved using the same process.
2. Digital Recognition
2.1 training set
Recognize handwritten numbers are used to study a classic neural network project. Handwritten digit recognition # # standard library is MINIST handwritten digital data set .
MINST data sets on LeCun's blog ( http://yann.lecun.com/exdb/mnist/ ). In this book the way with the code downloaded from the Web site data collection, and the downloaded data into unified data word is typically exists .pkl file. This operation requires a lot of time to use libraries Python, and the format operation of storing data, programming can be seen that there is high demand.