AI and deep learning调研报告

电子科技大学 格拉斯哥学院 2017届蔡昊宸
This article is an introduction to artificial intelligence (AI) and the interpretation and utilization of neural networks in the important part of today’s artificial intelligence machine learning. Artificial intelligence is presented by human entities, which are often considered computers. Artificial intelligence systems are now often used in economics, medicine, engineering, and military, as well as in traditional home computer software applications, traditional strategy games such as computer chess and other video games. We try to explain the brief concept of artificial intelligence and its application in various fields. It removes the concepts of computing and traditional categories. It includes a variety of advanced systems such as neural networks, fuzzy systems and evolutionary computing. AI is used for typical problems such as pattern recognition, natural language processing, and more. The system operates as an artificial brain worldwide. Intelligence involves mechanisms, and artificial intelligence research has found ways to make computers perform some of them rather than other computers. Computer programs can give very impressive performance on these tasks if the task requires only a mechanism that is well understood today. These programs should be considered “somewhat smart.” It is related to similar tasks using computers to understand human intelligence. We can understand how to make a machine solve a problem by observing others or just by observing our own methods. Artificial intelligence researchers are free to use methods that people have not observed, or involve more computational methods than people.
Advanced artificial intelligence has been applied to computer systems and programs that are capable of performing more complex tasks than basic programming, albeit still far from the realm of thought. The most important research areas in this field include practical areas such as information processing, pattern recognition and medical diagnosis. The latest artificial intelligence can understand the voice and the gaming computer being created, and even beat the best human players. Further research on information processing includes procedures that enable computers to understand written or verbal information to generate abstracts, answer specific questions, and redistribute information. Common examples of artificial intelligence in the contemporary world can be found in game activities such as puzzles, chess, checkers and cards. Voice technology is designed for use in artificial intelligence systems that not only include emerging technologies, but are also used by most of the world’s largest companies. Artificial intelligence is used in computer vision for functions such as image guided surgery, image recognition, facial recognition, and virtual reality. Speech recognition is now widely used for a variety of functions, such as customer service and banking, as well as dictation software and helping people with disabilities. Deep Blue is a program developed by IBM for playing chess. This show defeated world chess champion Gary Kasparov. Proverbs are artificial intelligence systems that solve crossword problems more effectively than humans. Watson is a answering machine that can make different TV game shows. Infotame is a text analytics search feature originally designed by the Russian KGB to intercept communications.
Machine learning is the core concept of artificial intelligence. The two are usually used interchangeably and have the same meaning. Suppose that instead of trying to design a comprehensive and straightforward instruction set that the system should basically follow to achieve their goals, a few examples of relevant data are provided so that they can develop their own identification systems and operational rules. Examples in this context are features of data and information that can be used to develop thousands of images through different sources of knowledge in order to identify different settings and words in handwriting. Alternatively, they can perform better identification work than systems that do not provide data on conventional doses.
Deep learning will become an important variable in the future intelligent growth equation, as it is now. Deep learning can have a huge impact, whether it’s for language or visual or natural language processing. Although deep learning has not proven to be a powerful tool for machine reasoning, it is a skill they must master to enhance many AI applications. Deep learning has successfully solved how to define problem types, which have a large amount of tagged data, which is beneficial to machine perception and classification problems, rather than true reasoning problems.
A neural network is a set of algorithms that are loosely modeled after the human brain to identify patterns. They interpret sensory data through a machine that perceives, marks or clusters the original input. The patterns they recognize are digital and are included in the vector. All real-world data, whether images, sounds, texts or time series, must be translated.
Neural networks help us cluster and classify. You can think of them as clustering and sorting layers on top of storing and managing data. They help to group unlabeled data based on similarities between the sample inputs and classify the data as it is trained with the tagged data set. (Neural networks can also extract features of other algorithms for clustering and classification; therefore you can think of deep neural networks as components of larger machine learning applications involving reinforcement learning, classification, and regression algorithms.)
All classification tasks depend on the marked data set; that is, humans must pass their knowledge to the data set so that the neurons learn the correlation between the label and the data. This is called supervised learning.
Detect faces, identify people in images, recognize facial expressions (angry, happy)
Identify objects in the image (parking signs, pedestrians, lane markings…)
Identify gestures in the video to detect speech, identify speakers, transcribe speech into text, and recognize emotions in speech
Classify text as spam (in email) or fraudulent (in insurance claims); Identify emotions in text (customer feedback)
Any tag that humans can generate, any results you care about, and any results related to the data can be used to train the neural network.
By classification, deep learning can establish a correlation between, for example, the pixels in the image and the names of the people. You can call it static prediction. For the same reason, exposure to enough correct data, deep learning can establish a correlation between current and future events. It can return between the past and the future. In a sense, future events are like labels. Deep learning doesn’t necessarily care about time, or things haven’t happened yet. Given a time series, deep learning can read a string of numbers and predict the next most likely number.
Hardware failure (data center, manufacturing, transportation)
Health failures (strokes, heart attacks based on important statistics and data from wearable devices)
Customer churn (predicting the likelihood of customer departure based on web activity and metadata)
Employee turnover (ibid, but employees)
The more we can predict, the more we can prevent and prevent it. As you can see, through neural networks, we are heading towards an unexpectedly reduced world. Not a zero accident, only a slight decrease. We are also moving towards a world of intelligent agents that combine neural networks with other algorithms such as reinforcement learning to achieve goals.
Deep learning is the name we use for “stacking neural networks”; that is, the network consists of several layers.


These layers consist of nodes. The node is just where the calculation takes place, loosely patterned on the neurons in the human brain, and emitted when it encounters enough stimulus. A node combines input from data with a set of coefficients or weights that amplify or suppress the input, thereby assigning the importance of the input to the task the algorithm is trying to learn. (For example, which input is most useful is to classify the data without errors?) These input weight products are summed, and the sum is passed through the node’s so-called activation function to determine if and to what extent the signal is further affected by the network. The end result, such as classification behavior.
The difference between a degree learning network and the more common single hidden layer neural networks lies in their depth; that is, the number of node layers that data passes through in the multi-step pattern recognition process.
Earlier versions of the neural network (such as the first perceptron) were shallow, consisting of an input layer and an output layer with at most one hidden layer. More than three layers (including input and output) are eligible for “depth” learning. Being so profound is a strictly defined technical term that means more than one hidden layer.
In a deep learning network, each layer of nodes trains a different set of features based on the output of the previous layer. The further the distance into the neural network, the more complex the features of the nodes are identified because they aggregate and recombine the features of the previous layer.
This is called a feature hierarchy, which is a hierarchy that adds complexity and abstraction. It enables deep learning networks to process large, high-dimensional data sets with billions of parameters through nonlinear functions.

Most importantly, these networks are able to discover potential structures in unlabeled unstructured data, which is the vast majority of data in the world. Another word for unstructured data is the original media; that is, images, text, video, and recording. Therefore, one of the most problematic problems of deep learning is to process and cluster the world’s original, unlabeled media, identifying similarities and anomalies in human data that are not organized or ever named in a relational database.
For example, deep learning can take a million images and cluster them according to their similarity: a cat in a corner, an icebreaker in another corner, and a third photo of a grandmother. This is the basis of the so-called smart photo album.
Now apply the same idea to other data types: deep learning may aggregate raw text, such as emails or news articles. E-mails full of anger complaints may gather in one corner of the vector space, and satisfied customers or spambot messages may gather in other corners. This is the basis for various messaging filters that can be used for customer relationship management (CRM). The same applies to voice messages. For time series, the data may revolve around normal/healthy behavior and abnormal/dangerous behavior. If time series data is generated by a smartphone, it will provide insight into the health and habits of the user; if it is generated by car accessories, it may be used to prevent catastrophic failure.
Unlike most traditional machine learning algorithms, deep learning networks perform automatic feature extraction without human intervention. Given that feature extraction is a task that can be done by the data scientist team for many years, deep learning is a way to circumvent the bottleneck of limited experts. It enhances the capabilities of small data science teams that are inherently incapable of scaling.
When training unlabeled data, each node layer in the deep network automatically learns the features by iteratively attempting to reconstruct the input from which the samples are taken, attempting to minimize the difference between the network guess and the probability distribution of the input data itself. For example, a restricted Boltzmann machine creates a so-called reconstruction in this way.
In the process, these networks learn to recognize the correlation between certain related features and the best results - they establish a connection between the signature signals and the content they represent, whether they are completely reconstructed or tagged.
Then, a deep learning network trained on tagged data can be applied to unstructured data, enabling it to access more input than machine learning networks. This is a way to improve performance: the more data the network can train, the more accurate it is. (A bad algorithm for training large amounts of data can outperform an excellent algorithm that is rarely trained.) The ability of deep learning to process and learn a large amount of unlabeled data gives it a distinct advantage over previous algorithms.