深度学习的三重境界：3LevelsinDeepLearning

深度学习(Deep Learning)是机器学习（Machine Learning）的一个分支。深度学习以人工神经网络（Artificial neural networks）为基础，进行监督、半监督、和无监督的学习。

Deep learning is a branch of Machine Learning. Based on Artificial neural networks, deep learning can be supervised,semi-supervised or unsupervised.

机器学习是人工智能的一个分支。机器学习根据数据的结构和推断，研究计算机系统中在没有明确指令情况下的算法和统计模型。它根据样本数据，也叫训练数据，建立数学模型，然后进行预测和判断。机器学习用途广泛，可以用于识别图片、过滤邮件等方面。

Machine learning is a subset of artificial intelligence.It is the study of algorithms and statistical models that computer programs use to perform specific tasks without human instructions. Machine learning build mathematical and statistical models based on sample data, also called training data, to forecast certain tasks. Machine Learning algorithms have many applications from computer vision to email filtering.

大量数据的出现，促进了机器学习和深度学习的发展。现在，数据量平均每18个月就增长一倍,处理数据的成本平均每两年就下降一半。

Large volume data help the development of machine learning and deep learning. Average speaking, data double every 18 months and processing cost halves every 2 years in nowadays.

第一层：昨夜西风凋碧树，独上高楼，望尽天涯路。

Level 1：Westerly winds withered trees up last night. Climbing up the stairs and being lonely on the loft, I overlooked the endless distance.

深度学习的基础是统计学。统计学最常见的模型是线性模型y=a+bx。这个模型可以用来刻画明显的线性关系，例如速度和距离，产品销量和企业利润等。但对于更复杂的关系，例如太极图中的黑白分布，就不能用线性模型来描述。而深度学习在这方面就更胜一筹。

The foundation of deep learning is statistics. The most popular statistical model is linear model: y=a+bx. This model can describe linear relationship such as velocity and distance, sales quantities and a company's profits. However, for more complex relationship,such as the black and white in Taiqi, linear models do not work. Deep learningcan solve this problem better.

                         

在Feifei li(2009)等人的文章“图片网络：海量多层次图片数据库”以及Krizhevsky （2012）等人的文章：“卷积深度神经网络网络图片分类”中，都使用了深度学习的方法对大量的网络图片进行归类。

Both the paper “ImageNet: A Large-Scale Hierarchical Image Database”by Feifei li（2009）and the paper “ImageNet Classification with Deep Convolutional Neural Networks”by Krizhevsky（2012）use deep learning methods to group images online.

神经网络是深度学习的基础，深度学习模型也被称为深度神经网络。所谓“深度”主要是指神经网络中隐含的层级。传统的神经网络隐藏2-3层。而深度网络可以隐含150层。

Most deep learning models use neural network architectures. Therefore, deep learning models are often referred to asdeep neural networks. “Deep” means the number of layers in the neural network. Classical neural networks contain 2-3 hidden layers. There could be 150 hidden layers in the deep networks.

第二层：衣带渐宽终不悔，为伊消得人憔悴。

Level 2: The dress takes to loosen graduallyand I am more and more emaciated, No regretful plying at all, I am rather forher only distressed as I did

机器视觉（Machine vision）是自动检测、处理图片的新技术。这不但可以将人从工作中解放出来，也可以防止肉眼检测中出现的错误。

Machine Vision is the new technology to provide imagine based automatic inspection. These new methods not only free people from work, but also prevent errors from human eyes.

胶囊神经网络（Capsule Neural Network）同样是一种人工神经网络（Artificial neural network）的机器学习系统，它更好的刻画了多层次的关系。这个方法尽量去模仿生物的神经系统。

Capsule neural network (CapsNet) is also a machinelearning approach based on artificial neural network. CapsNet models better hierarchical relationships. This method is more closely to mimic biological neural organization.

其他深度学习的方法包括自然语言识别（Natural language process）、时间序列(Time series)、循环神经网络(Recurrent neural network)。

Other deep learning methodsinclude Natural language processing (NLP), time series (TS) and recurrentneural network(RNN).

第三层：众里寻他千百度，蓦然回首，那人却在灯火阑珊处。

Level 3：Looking for him in the crowd, suddenly looking back, theman is in the lamp languishing place.

Tensor Flow最初由谷歌大脑团队开发，用于谷歌的生产和研究，于2015年11月发布。它是一个开源软件库，用于各种感知和语言理解任务的机器学习。在谷歌的商业产品如语音识别、gmail、相册和搜索中，都用到了TensorFlow.

TensorFlow is developed by Google Brain. It was published in Nov,2015 and used for Google’s production and research. TensorFlowis used in Google’s products such as gmail, photo album and search.

Keras是由Python语言编写而成的开源神经网络库。它可以作为TensorFlow,Microsoft-CNTK和Theano的高阶应用程序结构进行深度学习模型的设计、调试、评估、应用以及可视化。

Keras is an open-source neural network library.It’s written by Python. It is capable of running on top of TensorFlow, Microsoft Cognitive Toolkit,Theano.

Pytorch是由脸书公司人工智能团队开发的一款产品。它是根据Torch开发的开源机器学习库。PyTorch前端除了可以是python以外，也可以是C++。Uber的Pyro概率程序软件在后端就用到了PyTorch。

PyTorch is developed by Facebook’s artificial intelligence research group. It’s a free and open-source software. Uber’s Pyro probabilistic programming language use PyTorch as the backend.