AI and deep learning

The key of AI is machine learning, machine learning breakthrough is deep learning, artificial neural network.

In 1956, in the Dartmouth Conference (Dartmouth conferences), computer scientists first introduced the term "AI", the AI was born, and in subsequent days AI became the "fantasy object" of the lab. Decades later, people's view of AI is changing, sometimes think AI is a harbinger, is the key to the future of human civilization, sometimes think it is just technical waste, but a rash concept, ambition too big, doomed to failure. Frankly speaking, Ai still has both of these characteristics until 2012.

In the past few years, the outbreak of AI has been very rapid since 2015. The rapid development is largely due to the widespread popularity of GPUs, which make parallel processing faster, cheaper, and more powerful. Another reason is the unlimited expansion of real storage capacity, large-scale data generation, compared to slices, text, transactions, map data information.

AI: Let the machine show human intelligence

Back in the summer of 1956, at the time of the meeting, the AI pioneer dreamed of building a complex machine (which was driven by the computer that had just appeared) and then giving the machine a characteristic of human intelligence.

This concept is what we call "strong AI", which is to build an awesome machine that has all the human senses, and even beyond human perception, to think like a human being. In the film we often see this kind of machine, such as C-3PO, Terminator.

There is also a concept of "weak artificial intelligence (Narrow AI)". In short, "weak AI" can accomplish certain tasks like human beings, possibly better than humans, for example, Pinterest service uses AI to classify images, and Facebook uses AI to recognize faces, which is "weak AI".

The above example is a case of "weak artificial intelligence", which already embodies some of the characteristics of human intelligence. How is it achieved? Where does this intelligence come from? With the problem we understand deeply, and come to the next circle, it is machine learning.

Machine learning: A path to an AI target

In general, machine learning is to use algorithms to truly parse data, to learn constantly, and then to make judgments and predictions about what is happening in the world. Instead of writing software, defining special instruction sets, and then getting the program to do special tasks, researchers will "train" machines with a lot of data and algorithms to let machines learn how to perform tasks.

The concept of machine learning was proposed by early AI researchers, and in the past few years there have been many algorithmic approaches, including decision tree learning, inductive logic programming, cluster analysis (clustering), reinforcement learning, Bayesian networks, and so on. As we all know, no one really achieves the ultimate goal of "strong AI", using the early machine learning method, we even the goal of "weak artificial intelligence" is far from being achieved.

For many years, the best application case for machine learning has been "computer vision", where researchers still need to write a lot of code to accomplish their tasks to achieve computer vision. The researchers wrote the classifier manually, such as The edge detection filter, so that the program can determine where the object starts, where it ends, and shape detection can determine if the object has 8 edges; The classifier can recognize the character "S-t-o-p". By hand-written groupings, researchers can develop algorithms to identify meaningful images and then learn to judge that it is not a stop sign.

This method can be used, but not very good. If it is in foggy weather, when the visibility of the sign is low, or if a tree blocks a part of the sign, its ability to recognize will decline. Until recently, computer vision and image detection techniques were far from the human capability because it was so prone to error.

Deep Learning: The technology to realize machine learning

"Artificial Neural networks (Artificial neural Networks)" is another algorithmic approach, which is also proposed by early machine learning experts and has been in existence for decades. The idea of neural networks (neural Networks) stems from our understanding of the human brain-the mutual connection of neurons. There are also differences between the neurons of the human brain that are connected by a specific physical distance, the artificial neural network has a separate layer, the connection, and the direction of data transmission.

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

Each neuron weighs the input information, determines its weight, and makes sense of how it relates to the task it is performing, such as how correct or incorrect it is. The final result is decided by ownership. Take the stop sign as an example, we will cut off the logo image, let the neuron detection, such as its octagonal shape, red, distinctive characters, traffic sign size, gestures and so on.

The task of a neural network is to give a conclusion whether it is a stop sign or not. The neural network will give a "probability vector", which relies on the inferred and weighted weights. In this case, the system has 86% confidence that the picture is a stop sign, 7% confidence determines it is the speed limit sign, there is 5% confidence that it is a kite stuck in a tree, and so on. The network architecture then tells the neural network if it's judged correctly.

Even if it's just such a simple thing to be ahead of itself, not long ago, the AI research community is still avoiding neural networks. Neural networks have existed in the early stages of AI development, but it has not formed much "intelligence". The problem is that even basic neural networks, which have a high demand for computational capacity, cannot be a practical approach. Nonetheless, a handful of research teams, such as the Geoffrey Hinton of the University of Toronto, led the team to put algorithms in parallel to supercomputers and validate their concepts until the GPU began to be widely adopted before we really saw hope.

Back to the example of identifying stop signs, if we train the network, train the network with a lot of wrong answers, and adjust the network, the results will be better. What the researchers need to do is train them to collect tens of thousands of or even millions of pictures, until the weight of the artificial neuron input is highly accurate, so that every judgment is correct-whether it's fog or fog, it's sunny or rainy. At this point the neural network can "teach" themselves, to find out what the stop sign is exactly, it can also identify Facebook face images, can identify the cat-Wunda (Andrew Ng) in 2012, Google did the thing is to let the neural network to identify the cat.

Wunda's breakthrough is to make the neural network extremely large, increasing the number of layers and neurons, allowing the system to run large amounts of data and train it. Wunda's project calls images from 10 million YouTube videos, and he really gives deep learning a "depth".

Today, in some scenarios, machines trained in deep learning techniques are better at identifying images than humans, such as identifying cats, identifying cancer cells in the blood, and identifying tumors in MRI scans. Google Alphago learning go, and it itself and constantly go and learn from it.

With deep learning AI, the future is bright.

With deep learning, machine learning has many practical applications, and it expands the overall scope of AI. Deep learning splits tasks so that various types of machine assistance become possible. Driverless cars, better preventative treatments, better film recommendations either have appeared, or even appear. AI is both present and future. With the help of deep learning, maybe one day AI will reach the level that science fiction describes, which is what we've been waiting for for a long time.

Http://newseed.pedaily.cn/201610/201610091326716.shtml

AI and deep learning