Deep learning - Acquaintance deep learning - Notes

 Traditional machine learning and deep learning

　　Despite the depth of learning as the traditional machine learning on the ability to learn, but there are still deep learning machine learning shadow, deep learning theory is inseparable from the traditional machine learning. Such as a linear model, the convolution kernel, RELU activation function, SoftMax, all of neural network learning in depth. Following a brief comparison of traditional machine learning and deep learning.

Traditional machine learning process of establishing

Data preprocessing -----------> feature extraction -----------> selection classifier
Normalized Dimensionality reduction De-noising	Artificial design patterns Image: SIF, LBP, Fisher, Gabor, Hog Language: MFCC, wavelet, Word2vec	SVM, decision trees, random forest, Bayesian networks, linear regression, clustering

Deep learning process of establishing

Data Preparation -----------> Design Model -----------> Training
Data, label	CNN, CNN + RNN RNN	Structural adjustment loss function training parameters

Depth learning applications Features

advantage:

　　Learning ability, covering a wide range of adaptability, portability (such as migration study)

Disadvantages:

　　Computationally intensive, poor portability, high hardware requirements, complex model design, there are likely to be "hack" - it is possible to recognize certain patterns into other objects, longer than the calculation, calculation is weaker than

Relatively deep learning framework

frame	Language	Documentation	CNN compatible	RNN compatible	Difficulty to get started	speed	Parallel Support	Keras compatible	Support Team
Theano	Python/C++	++	++	++	+	++	+	+	University of Montreal
Tensor Flow	Python	+++	+++	++	+++	++	++	+	Google
Torch/Pytorch	Lua, Python	+	+++	++	++	+++	++		Facebook
Caffe	C++	+	++		+	+	+		Jia Yang Qing, Berkeley, California
MXNet	Python,R,Julia	++	++	+	++	++	+++	+	Li Mu, Chen Tianqi, etc., Amazon
Neon	Python	+	++	+	+	++	+		Intel
CNT	C++	+	++	+++	+	++	+		Microsoft

TensorFlow

Rich document, suitable for beginners, are easy to install Google support, long-term effective, automatic derivation, just focus on model design, Keras support to facilitate the rapid development.

Specific applications TensorFlow can look at a number of Chinese sites like TensorFlow Chinese community http://www.tensorfly.cn/tfdoc/how_tos/overview.html , also c language Chinese network http://c.biancheng.net/ tensorflow / , of course, you can learn the relevant documents in English.

Learn the basics of deep learning

　　Artificial neural network called neural network is a mathematical model draws on the principle of biological neural networks formed. A neural network is one of many machine learning algorithms, both can be used for supervised tasks, such as classification, visual recognition, etc., it can also be used unsupervised tasks. At the same time it is capable of handling complex non-linear problem, its basic structure is a neuronal, as shown below: This is a simple one only neurons

 

 Wherein, x1, x2, x3 represents the input, intermediate portions of neurons, and the last hw, b (x) is the output neuron. The whole process can be understood as an input -> Processing -> Output.

　　A plurality of neurons of the neural network is: It consists of an input layer, a hidden layer and output layer, as shown iris data set is the results after training, of course, this is a theft FIG other blog of:

Convolution kernel is the basic operator of image processing, neural network convolution convolution kernel applied to the neural network, you can find out about the process from the convolution of the figure below:

 The next step is learning to start learning the depth of some of the systems.