Liu Youjuan drawing When I'm alone, I feel lonely. What should I do? The "Microsoft Ice" launched by Microsoft Research Asia may be able to chat with you like a girlfriend. The 3.0 version of "Little Ice" not only has such functions as "appraisal value" and "select and match", but also has powerful visual recognition ability based on deep learning technology. After seeing a picture, it can give a humanized reply based on emotions, and the speed of seconds is shortened to 250 milliseconds. Not only "Microsoft Little Ice" and Go Master "Alfa Dog", from Internet search to language translation, and even identifying genes with risk of autism... All fields that need to predict unknown information from large amounts of data are in-depth Learn where you can flex your muscles. So, what is deep learning technology? How will it change human life? In 2011, researchers from a Google lab took 10 million static images from video sites and "fed" them to Google's brain. The goal was to find recurring patterns. Three days later, the Google brain discovered "cats" from these pictures without human help. The Google brain is a large neural network model that uses deep learning technology and consists of 1,000 computers. This incident caused a sensation in the science and technology industry at that time and was considered as a milestone in deep learning revival. The so-called deep learning is a neural network composed of multiple layers of neurons to achieve the machine learning function. These multi-layered computer networks, like the human brain, collect information and generate corresponding behavior based on the collected information. Traditional machine learning methods can only mine a simple linear relationship, such as 1+1 is equal to 2. However, Daqian is not what this simple relationship can describe, such as the relationship between income and age, gender, occupation, and education. The appearance of deep learning has changed this situation. Its inspiration comes from imitating human brain neural networks. Scientists have discovered that the human cerebral cortex does not directly extract features from the data transmitted by the retina. Instead, it filters the received stimuli through a complex network model. This hierarchical structure greatly reduces the amount of data processed by the vision system. And eventually retained useful information. In the 1960s, when biologists studied the cat's cerebral cortex, they discovered that their unique network structure could effectively reduce the complexity of the feedback neural network, and then proposed the "convolutional neural network." The deep learning program written using this kind of network structure has strong adaptability and becomes a breakthrough of artificial intelligence. Simply speaking, deep learning technology is a simulation of the human brain and can therefore complete the functions of many human brains. The best known is the visual function. Our camera can see the world like the eyes, but it can't understand the world like the brain. Deep learning just fills this short board. With deep learning, Baidu maps can accurately identify the types of objects in photos and automatically classify or search for photos. With deep learning, we can easily pay for the face. With deep learning, special machines can detect the whereabouts of all personnel and vehicles in a certain space and promptly report suspicious and dangerous events. At the same time, deep learning technology has also been widely used in speech recognition. With the help of deep learning, computers have more and more powerful voice recognition capabilities, which may change the human-machine interaction mode that is still dominated by the keyboard. The combination of deep learning and enhanced learning is profoundly changing the field of robotics. The so-called enhanced learning refers to the robot's self-study of better strategies through rewards and punishments obtained through interaction with the environment. The "Alfa dog" that has attracted people's attention some time ago is the product of enhanced learning. It has found a better strategy for playing chess by playing chess with a chess player or winning or losing with oneself. However, creating a powerful neural network requires more layers of processing. Due to hardware limitations, only 2 to 3 neural layers can be made early. So what makes deep learning beyond? Obviously, the improvement of high-performance computing capabilities is a major boost. The rapid development of GPUs, supercomputers, and cloud computing over the years has made deep learning stand out. In 2011, Google's brain used 1,000 machines and 16,000 CPUs to process about 1 billion neurons in a deep learning model. And now, we can already do the same calculation on several GPUs. “In-depth learning is also helped by big data, just like the rockets have fuel.†Dr. Pan Zheng from the Department of Automation of Tsinghua University, a computer vision engineer from Geling Deep-pupil, said that deep learning technology is based on a large number of examples, just like children collect reality. The world's information is the same. Moreover, the more data you feed, the smarter it will be, and it will not "digest." Because of the indispensability of big data, the best basic level of deep learning is IT giants with a lot of data, such as Google, Microsoft, and Baidu. Nowadays, deep learning technology overcomes traditional machine learning methods in the fields of speech recognition, computer vision, and language translation, and even surpasses human recognition in face verification and image classification. Experts predict that in a few years, our cell phone in the pocket can run a neural network that is as complex as the human brain. However, as far as the current trend is concerned, deep learning technology still cannot replace the “people who sit in the background monitoring roomâ€. For example, if you and your friends are rushing to checkout after having a meal in a restaurant, it is difficult for the smart camera to judge whether it is fighting or something. It can be seen that logical judgments and emotional choices are obstacles that in-depth learning is insurmountable. A bad system can be identified at a glance Focused on computer vision and artificial intelligence technology company Geling Shen pupil, will be based on deep learning technology research and development of intelligent identification system, applied to the field of bank security monitoring. Taking into account the traditional optical lens will lose the "depth" dimension in the recognition of images, Geling deep pupil specifically developed a set of three-dimensional sensors for bank safety supervision. Behind it, a set of algorithm models trained by reward and punishment mechanisms can actively identify abnormalities. “Seeing someone approaching a human ATM, instead of being next to it, at this time to identify his trajectory and determine whether his behavior is normal involves deep learning.†If the system recognizes an anomaly, it will be pushed to the background monitor. In order to teach the machine to be accurate, hundreds of thousands of images need to be provided. He Yongfei pointed out that given the smart identification system a side face or a faceless full body shot, it can also quickly lock the target with more than 99% accuracy. The premise is to build a sample database of 6,000 to 15,000. "Once the sample reaches a million, the accuracy may decrease by 20% or more." Other household electric appliance gree , https://www.greegroups.com