A diagram shows the difference between AI, machine learning, and deep learning

AI (artificial intelligence) is the future, it is science fiction, is part of our daily life. All judgment is right, just to see what you talked about AI.

For example, when Google DeepMind AlphaGo program development professional go beat South Korea master Lee Se - dol, media in the description of the DeepMind victory in the AI, machine learning, deep learning and other terms. AlphaGo beat Lee Se - dol, these three techniques are set well, but they are not the same thing.

To find out their relationship, the most intuitive expression is concentric circles, the first is the idea, then the machine learning, when the machine learning boom in depth study, today's AI outbreak is driven by deep learning.

From decay to prosperity.

In 1956, at the Dartmouth Conferences, computer scientists first coined the term "AI", which was born, and in the days that followed, AI became the laboratory's "fantasy object". As the decades passed, the people of AI opinion changing, sometimes think that AI is warning, is the key to the future of human civilization, sometimes think of it as junk technology, is the concept of a rash, excessive ambition, is doomed to fail. Frankly, the AI still has both of these characteristics until 2012.

In the past few years, AI has exploded, and it has been developing rapidly since 2015. The rapid development is largely due to the widespread popularity of GPU, which makes parallel processing faster, cheaper and more powerful. Another reason is that the actual storage capacity is unlimited, and the data is generated on a large scale, such as images, text, transactions, and map data.

AI: let the machine show human intelligence.

Back in the summer of 1956, at the time of the meeting, the AI pioneers dream is to build a complex machine (for emerging computer driver), and then let the machine shows the characteristics of human intelligence.

This concept is what we call "strong artificial intelligence (AI) General", which is to build a great machine, let it all with human perception, you can even go beyond human perception, it can think like a man. In movies, we often see this kind of machine, like c-3po, terminator.

Another concept is "Narrow AI". Simply speaking, the "weak" artificial intelligence can complete some specific task like humans, may do better than humans, for example, Pinterest services with AI to image classification, Facebook with AI face, this is the "weak" artificial intelligence.

The examples above are examples of the actual use of "weak artificial intelligence", which has already demonstrated some of the characteristics of human intelligence. How? Where does this intelligence come from? With the problem we understand, we go to the next circle, which is machine learning.

Machine learning: a path to the AI goal.

In general, machine learning is to use algorithms to really parse data, to learn, and to make judgments and predictions about what's going on in the world. At this point, the researchers don't have to write software, determine the specific instruction set, and then let the program for special tasks, on the contrary, the researchers will use large amounts of data and algorithm "training" machine, let the machine learn how to perform a task.

Machine learning is the concept of early AI researchers put forward, in the past few years, there appear a lot of algorithm method, machine learning decision tree learning, inductive logic programming, including cluster analysis (Clustering), reinforcement learning, bayesian networks, etc. As we all know, no one really achieves the ultimate goal of "strong artificial intelligence", and the goal of "weak artificial intelligence" is far from being achieved by adopting early machine learning methods.

In the past many years, the best use case for machine learning was "computer vision," and to achieve computer vision, researchers still had to write a lot of code manually to complete the task. Researchers manually write classifiers, such as edge detection filters, to determine where an object starts and ends. Shape detection can determine whether an object has 8 edges; The classifier can recognize the character "s-t-o-p". By manually writing the pacers, the researchers can develop algorithms to identify meaningful images and then learn to determine that it is not a stop sign.

It works, but it's not very good. If it is foggy, when visibility is low, or a tree blocks part of the sign, its ability to recognize is reduced. Until recently, computer vision and image-detection techniques were far from human capabilities, because they were too prone to error.

Deep learning: the technology to realize machine learning.

Artificial Neural Networks (Artificial Neural Networks) is another algorithmic method, which has been proposed by early machine learning experts for decades. The idea of Neural Networks stems from our understanding of the human brain -- the connection between neurons. There are differences, too. Neurons in the human brain are wired to specific physical distances, and artificial neural networks have separate layers, connections, and data transmission directions.

For example, you might draw a picture, cut it into chunks, and implant it into the first layer of the neural network. The first layer of independent neurons sends the data to the second layer, and the second layer has its own mission, which continues until the last layer, and produces the final result.

Each neuron will weigh the input, determine the weight, and figure out how it relates to the task performed, such as how correct or incorrect it is. The final result is determined by the weight of ownership. Take the stop sign as an example, we will cut the stop sign image and let the neuron detect, such as its octagonal shape, red, distinctive characters, traffic sign size, gesture, etc.

The task of the neural network is to give a conclusion: is it a stop sign? The neural network will give a "probability vector", which depends on educated guesses and weights. In this case, the system had 86% confidence that the picture was a stop sign, 7% confidence that it was a speed limit sign, 5% confidence that it was a kite stuck in a tree, and so on. The network architecture then tells the neural network whether it is correct or not.

Even if it's just that simple, it's pretty futuristic, and not long ago, the AI community was avoiding neural networks. In the early days of AI development, neural networks existed, but it did not form much "intelligence". The problem is that even the basic neural network, which has high computational requirements, cannot be a practical approach. Still, there are a few research groups, such as the university of Toronto led by Geoffrey Hinton's team, they will parallel algorithm in the super computer, verify their concepts, we didn't really see until GPU began to widespread adoption of hope.

Back to the example of identifying stop signs, if we train the network, use a lot of wrong answers to train the network, adjust the network, the result will be better. , researchers need to do is to train them to collect tens of thousands of copies, or even millions of images, until artificial neuron input the weight of high precision, let every judgment right - whether it's fog or mist, is it sunny or rain are not affected. At this point, the neural network can "teach" itself and find out what the stop sign is. It can also identify Facebook's face images and identify the cat -- Andrew Ng's 2012 Google job -- to get the neural network to recognize cats.

Mr Ng's breakthrough was to make the neural network infinitely large, increasing the number of layers and the number of neurons, allowing the system to run large amounts of data and train it. Mr Ng's project, which calls images from 10m YouTube video, really gives depth to deep learning.

Today, in some scenarios, through deep learning machine technical training in identifying the image is better than the human, such as cats, identify the characteristics of the cancer cells in the blood, recognizing tumors in MRI scanning images. Google AlphaGo learns the game of go and learns from it itself and from it.

The future is bright with deep learning AI.

With deep learning, machine learning has a lot of practical applications. It also expands the overall range of AI. Deep learning splits the task, making it possible for various types of machine assistance. Driverless cars, better preventive treatments, better movie recommendations have either already appeared or even appear. AI is both present and future. With the help of deep learning, maybe someday AI will reach the level of science fiction, which is exactly what we've been waiting for. You'll have your own c-3po, and you'll have your own terminator.