table of Contents
Part one. Artificial intelligence, machine learning and deep learning
First, artificial intelligence
1. The definition of artificial intelligence
2. Classification of Artificial Intelligence
1. The definition of machine learning
2. The machine learning classifier
3. common machine learning methods
1. The definition of deep learning
2. deep learning classification
The main application of deep learning
Part Two. In the English term control
Part one. Artificial intelligence, machine learning and deep learning
First, artificial intelligence
1. The definition of artificial intelligence
Research and development for simulation, extension and expansion of human intelligence an emerging discipline theories, methods, techniques and applications.
2. Classification of Artificial Intelligence
Weak AI: Think impossible to create a truly intelligent machine reasoning and problem solving, these machines look like intelligence, but does not really have a smart, and there will be no sense of ownership (pessimists)
Strong Artificial Intelligence: think it's possible to create a truly intelligent machine reasoning and problem solving, such a machine is considered to have a self-conscious (optimists)
Super artificial intelligence: machine intelligence thorough than humans (ultra-optimists)
Second, the machine learning
1. The definition of machine learning
Let the computer has the technical ability to think and learn like human beings in general. Specifically, technical laws obtained from known data and unknown data using the law of prediction
2. The machine learning classifier
Supervised learning (Supervised Learning): there is the case of teachers (environment), students (computer) to get there right or wrong instructions from the teacher (environment), the final answer of learning. Review with the school division. Briefly dataset label . For example: classification, regression
Unsupervised learning (Unsupervised Learning): no teacher (environment), the student (computer) self-learning process, the general use of some of the established criteria for evaluation. Standard self-assessment. In short dataset has no label . For example: clustering, dimensionality reduction
Reinforcement Learning (Reinforcement Learning): the absence of a teacher (environment), students (computer) method to answer questions of self-evaluation, emphasis was able to get feedback from the environment . Self-study self-assessment. Briefly large part of the data set does not label, a small portion of the label data set . For example: chess
3. common machine learning methods
Third, the depth of learning
1. The definition of deep learning
Generally refers to the depth of learning by training a multi-layer network structure were unknown data classification or regression
2. deep learning classification
Supervised learning method: feed-forward network depth before, convolution neural network, recurrent neural networks
Unsupervised learning methods: deep belief network, Boltzmann machine depth, the depth from the encoder to generate confrontation networks
3. Why use deep learning
For instance, we do a classification, this is my task to distinguish this is not a car. In reality, the car may have different angles of different size image, the car will have different scales, in different directions, and even different colors and the like. To the separation of this car, we must know the structure, and therefore to design some feature, such as a feature of our most popular car is the seat. To overcome this different scales and at different angles, the need to find his local maximum, then this maximum value above which a gradient histogram used to represent its structure, the gradient histogram can show his direction of the edge as well as his statistical distribution characteristics. Then put this feature on a machine learning model where, for example, a very common SVM, then training and classification.
That there is, factors affecting the performance of machine learning into two: the first is the model they fit, such as SVM which have chosen different parameters, of course, this can be approximated by a purpose-driven training as much as possible the best results; second factor is the feature that if this feature is not a good indication that the object, or the application is implemented, and that even though the latter model is better, it is difficult to get good performance. For example, here I do not have seat, only in color. But the color is completely impossible to determine this is not a car, so no matter how good you that the use of classifiers, it is impossible to classify the car out. Behind you even after convolution with a particularly good network, and more particularly, then, but before you enter this information is completely impossible to distinguish it, then it does not matter how good the back of the line. So in the tradition of this machine learning, it is very dependent on the feature of the design professionals.
But after learning the depth to it, people put feature extraction and classification combined together. These two parts are purpose-driven by way of training, for example, to get the end of this training method. This time it was, in fact, I do not know in the end I trained this feature is what kind of a structure, or what kind of physical meaning. I have not quite know, because he is completely process a large data trained. This is equivalent to become a black box, but this does, it is a purpose-driven to some extent, if the amount of data enough, he will be trained by the performance of the object (car) to some extent, the best feature. Therefore, the amount of data is large enough, and the network structure is relatively reasonable, then, the way we determine the effect will be more than the traditional manual design.
4. The depth profile learning
The main application of deep learning
(1) The image processing main applications:
- Image classification (object recognition): the classification or recognition of the entire image
- Object detection: detection of objects in the image of further object recognition
- Image segmentation: specific object in the image divided by Edge
- Image Regression: coordinate components of objects in an image prediction
(2) The main applications of speech recognition
- Speech recognition: the voice recognition to text
- Voiceprint identification: identify which person's voice
- Speech synthesis: the synthesis of specific human voice according to the text
(3) The main application field of natural language processing
- The language model: According to the forecast before the next word a word.
- Sentiment analysis: analysis of the text reflects the emotions (positive and negative to type, or positive and negative in attitude).
- Nerve Machine Translation: multilingual translation based on statistical language model.
- Abstract nerve Auto: Automatically generate a summary based on the text.
- Machine reading comprehension: answer questions by reading the text, complete multiple-choice questions or cloze.
- Natural language reasoning: According to the sentence (the premise) deduce another sentence (conclusion).
(4) integrated application
- Image Description: The description of the image given image sentences
- Visual Q & A: answer questions based on images or videos
- Image generation: generate an image based on the text description
- Video Generation: automatic generation of video based on the story
Part Two. In the English term control
❑ AI: Artificial Intelligence
❑ Computational Intelligence: Computational Intelligence
❑ aware intelligence: Perceptual Intelligence
❑ cognitive intelligence: Cognitive Intelligence
❑ Machine Learning: Machine Learning
❑ supervised learning: Supervised learning
❑ Unsupervised Learning: Unsupervised learning
❑ Enhanced Learning: Reinforcement Learning
❑ Neuron: Neuron
❑ Perceptron: Perceptron
❑ Neural Networks: Neural Networks
❑ back-propagation algorithm: Back Propagation, BP
❑ convolution neural network: Convolutional Neural Network , CNN
❑ deep learning: Deep Learning
❑ gradient disappears: Vanishing Gradient
Linear correction unit ❑: Rectified Linear Unit, ReLU
❑ deep belief network: Deep Belief Networks
❑ Boltzmann machine: Boltzmann Machines
❑ variational study: Variational Learning
❑ Classification: Classification
❑ recursion: Recursion
❑ deep belief networks: Deep Belief Network, DBN
❑ depth Boltzmann machine: Deep Boltzmann Machine, DBM
❑ depth from the encoder: Deep Auto-Encoder, DAE
❑ noise from the encoder: Denoising Auto-Encoder, D-AE
❑ Stacker from the encoder: Stacked Auto-Encoder, SAE
❑ generated against the network: Generative Adversarial Networks , the GAN
❑ nonparametric Bayesian Network: Non-parametric Bayesian Networks
❑ depth before feed-forward network: Deep Feedforward Neural Network, D-FNN
❑ convolution neural network: Convolutional Neural Network, CNN
❑ Recurrent Neural Networks: Recurrent Neural Network, RNN
❑ Network capsule: Capsule Net
❑ the depth of the forest: Deep Forest
❑ image classification (object recognition): Image Classification (Object Recognition)
❑ object detection: Object Detection
❑ image segmentation: Image Segmentation
❑ Return image: Image Regression
❑ Speech Recognition: Automatic Speech Recognition, ASR
❑ voiceprint identification: Voiceprint Recognition
❑ Speech synthesis: Speech Synthesis
❑ language models: Language Model
❑ sentiment analysis: Sentiment Analysis
❑ nerve Machine Translation: Neural Machine Translation, NMT
❑ nerve Automatic Summary: Neural Automatic Summarization
❑ machine reading comprehension: Machine Reading Comprehension, MRC
❑ natural language reasoning: Natural Language Inference, NLI
❑ contains the text: Text Entailment
