Artificial intelligence (AI) has entered the horizons of the general public, and we can see a lot of AI-related products in our daily life. Such as Siri, AI beauty, AI face change…

Although people have heard a lot about AI, most people don’t know much about it, and there are even some misunderstandings. This document does not involve any technical details, but only helps to understand the nature of artificial intelligence.

What is artificial intelligence?

Many people have some misconceptions about ARTIFICIAL intelligence:

1. Robots in movies are typical examples of artificial intelligence
2. Artificial intelligence seems to be omnipotent
3. Artificial intelligence will threaten the survival of mankind in the future
4. …

There are a lot of misconceptions about AI because people just read what people say, but they don’t understand the fundamentals of AI.

We compare it with traditional software and artificial intelligence. It’s easier to understand when you have a frame of reference.

Traditional software VS artificial intelligence

Traditional software

Traditional software is the basic logic of “if-then”. Human beings summarize some effective rules based on their own experience, and then let the computer run these rules automatically. Traditional software can never transcend the boundaries of human knowledge, because all the rules are made by humans.

To put it simply: traditional software is “rules-based”, requiring humans to set conditions and tell the computer what to do if it meets them.

This logic works well for simple problems where the rules are clear, the results are predictable, and the programmer is the software’s god.

However, real life is full of all kinds of complex problems, which are almost impossible to be solved by rules. For example, face recognition can not be solved by rules.

Artificial intelligence (ai)

Now artificial intelligence has developed many different branches, technology principles are also diverse, here only introduces the current hot deep learning.

Deep learning technology works in a completely different way from traditional software logic:

Machines summarize patterns from “specific” masses of data, generalize “specific knowledge,” and then apply this “knowledge” to real situations to solve real problems.

This is the essential logic of the development of artificial intelligence to the present stage. The knowledge summed up by artificial intelligence is not as intuitive and accurate as that expressed by traditional software. It’s more like human learning, abstract and hard to express.

Artificial intelligence is a tool

AI and hammers, cars, computers… It’s all the same. It’s all a tool.

Tools have to be used in order to be of value, and they are worthless if they stand on their own, like hammers in a toolbox that have no value unless wielded by someone.

Artificial intelligence is a tool that society is talking about because it greatly expands the capabilities of traditional software. There are many things that computers can’t do before, but now artificial intelligence can do them.

But no matter what happens, traditional software and artificial intelligence are tools, designed to solve real problems. That hasn’t changed.

Ai currently only solves specific problems

Terminator, Matrix… There are a lot of movies with crazy robots, which give people a sense that artificial intelligence is omnipotent.

The reality is that ai is still in the single-mission stage.

Single-task mode

Make a phone call on a landline, play games on a console, listen to music on an MP3 player, drive a car on navigation…

Multitasking mode

This stage is similar to smart phones. You can install many apps and do many things on one phone.

However, these abilities are independent of each other. After booking the air ticket on the travel App, you need to set the alarm clock by yourself using the alarm clock App, and finally you need to hail a taxi by yourself using the taxi App. Multitasking is just a superimposition of a single task, far from human intelligence.

Achieve mastery through a comprehensive

You are playing weiqi with your friend, and you find that your friend is in a bad mood. You could have won easily, but you deliberately lose to the opponent, and you keep praising him, because you don’t want to make him more depressed and irritable.

You use many different skills in this simple matter: emotion recognition, Go skills, communication, psychology…

But the famous AlphaGo would never do that. No matter what the other team’s situation is, even if losing the game will cost you your life, AlphaGo will win the match ruthlessly because it can do nothing but play Go!

Knowledge can only be integrated if it is formed into a network. ** For example, military knowledge can be applied to business, and biological knowledge can be applied to economics.

Know it, but don’t know why

There’s a big problem with today’s crude “induction” method of summarizing knowledge from vast amounts of data:

Don’t care why

Ponzi schemes take full advantage of this!

• It uses super-high returns to attract leeks, and then makes money for everyone who gets up early to participate;

• When onlookers see that all participants actually made money, they can simply conclude that historical experience makes sense.

• So more and more people are jealous and join in, until one day the cheater runs away.

When we apply logic to the story, we can conclude that the liar:

• Such high returns do not conform to market rules

• A sure bet? I don’t have to take the high risk of a high return, right? Doesn’t seem to make sense

• How could something so good happen to me? Something doesn’t seem right

Because current AI is based on inductive logic, it can also make rudimentary mistakes

And because of inductive logic, you have to rely on a lot of data.The more data, the more generalizable the experience.

The history of artificial intelligence

Between 1950 and 2017, there were some milestones in the field of artificial intelligence. In summary, it can be divided into three stages:

First Wave (non-intelligent conversational robots)

The 1950s and 1960s

In October 1950, Turing proposed the concept of artificial intelligence (AI), along with the Turing test to test AI.

Only a few years after the Turing test was proposed, people saw the “dawn” of a computer passing the Turing test.

1966 ELIZA, a psychotherapy robot, was born

People in those days thought highly of him, and some patients even liked to talk to robots. But his implementation logic is simple: a limited library of conversations, and when a patient says a certain key word, the robot responds.

The first wave didn’t use new technology. It used tricks to make computers look like real people. Computers were not intelligent.

Second Wave (Speech recognition)

The 1980s and 1990s

Speech recognition is one of the most iconic breakthroughs in the second wave. The core reason for the breakthrough is to give up the thinking of the symbolic school and to solve practical problems with statistical thinking.

In his book artificial Intelligence, Kai-fu Lee details this process and is one of the key figures involved.

The biggest breakthrough of the second wave was a change of thinking, abandoning the symbolic school of thought and using statistical thinking to solve problems.

Third wave (deep learning +Big data)

At the beginning of the 21st century

The year 2006 was a watershed in the history of deep learning. Jeffrey Hinton published “A Fast Learning Algorithm for Deep confidence Networks” and other important academic articles on deep learning were published in the same year, leading to several major breakthroughs in basic theory.

Two conditions are ripe for the third wave:

After 2000, the rapid development of the Internet industry has formed a massive amount of data. And the cost of data storage is falling fast. It makes it possible to store and analyze massive amounts of data.

The continuous maturation of Gpus provides necessary support for computing power, improves the availability of algorithms and reduces the cost of computing power.

After various conditions are mature, deep learning plays a powerful role. In the speech recognition, image recognition, NLP and other fields continue to refresh the record. For example, the error rate of speech recognition is only 6%, the accuracy rate of face recognition is higher than that of human beings, and BERT outperforms human beings in 11 performances… Phase.

The third wave is coming, mainly becauseBig dataAnd computing power, so deep learning can play a huge role, and AI has already outperformed humans, reaching the stage of “usable” rather than just scientific research.

Differences in the 3 waves of AI

1. The first two were academically driven; the third was driven by real business needs.
2. The first two were mostly about marketing, while the third was about business models.
3. In the first two booms, academics were trying to persuade governments and investors to invest money. In the third boom, investors took the initiative to invest money in academic and entrepreneurial projects in hot fields.
4. The first two booms raised more questions, and the third boom solved more.

The capabilities of artificial intelligence

Three levels of artificial intelligence

Weak artificial intelligence

Weak AI, also known as Narrow AI or Applied AI, refers to AI that focuses on and can only solve problems in a specific field.

Examples: AlphaGo, Siri, FaceID…

Strong artificial intelligence

Also known as Artificial General Intelligence or Full AI, it refers to Artificial Intelligence that can perform all human tasks.

Strong AI has the following capabilities:

• Ability to reason, use strategies, solve problems, and make decisions in the presence of uncertainty

• The ability of knowledge representation, including the ability of common sense knowledge representation

• Planning ability

• The ability to learn

• The ability to communicate in natural language

• The ability to combine these capabilities to achieve a given goal

Superartificial intelligence

Assuming that computer programs evolve to outsmart the world’s smartest and most talented humans, the resulting ARTIFICIAL intelligence system could be called superartificial intelligence.

We are currently at a stage of weak AI, strong AI is not there yet (not even close), and super AI is nowhere to be seen. So “domain-specific” is still a boundary that AI cannot cross.

What are the capabilities of AI?

To explain the boundaries of AI at a theoretical level, it takes Master Turing out of the picture. In the mid-1930s, Alan Turing was thinking of three questions:

1. Are there clear answers to all mathematical problems in the world?
2. If there is a clear answer, can the answer be calculated in a limited number of steps?
3. For a mathematical problem that can be solved in a finite number of steps, can there be a machine that keeps moving until at last, when the machine stops, the mathematical problem is solved?

Turing **** devised a method, later called the Turing machine. All computers today, including the new ones being designed all over the world, are no more than Turing machines in terms of their problem-solving power.

Through these three questions, Alan had drawn the line that would apply not only to today’s AI, but also to tomorrow’s AI.

1. There are many problems in the world, and only a small part of them are mathematical problems
2. Only a fraction of mathematical problems have solutions
3. Only part of the solvable problem can be solved by an ideal Turing machine
4. The latter part (the part that Turing machines can solve) is only part that today’s computers can solve
5. And the problems that AI can solve are only part of the problems that computers can solve.

In certain scenarios, the AI can work well, but in most scenarios, the AI is useless.

What jobs will be replaced by AI?

Kai-fu Lee proposed a judgment basis:

If a job takes less than five seconds to make a decision, it will most likely be replaced by ARTIFICIAL intelligence.

This kind of job has four characteristics:

1. The amount of information needed to make a decision is small
2. The decision-making process is not complicated, the logic is simple
3. Can be done alone, without collaboration
4. Repetitive work

Scientists have identified three skills that ai will have a hard time replacing:

1. Social intelligence (insight, negotiation skills, empathy…)
2. Creativity (originality, artistic aesthetics…)
3. Perception and manipulation (finger sensitivity, coordination, ability to deal with complex environments…)

Machine learning, machine learning | ML

What is the relationship between machine learning, artificial intelligence and deep learning?

In 1956, the concept of AI was introduced. In 1959, Arthur Samuel proposed the concept of machine learning:

Field of study that gives computers the ability to learn without being explicitly programmed.

Machine learning studies and builds special algorithms (rather than a specific algorithm) that allow computers to make predictions by learning from data themselves.

Therefore, machine learning is not a specific algorithm, but a general term for many algorithms.

Machine learning includes many different algorithms. Deep learning is one of them. Other methods include decision trees, clustering, bayes, etc.

Deep learning is inspired by the structure and function of the brain, the interconnection of many neurons. Artificial neural networks (ANN) are algorithms that simulate the biological structure of the brain.

Both machine learning and deep learning belong to artificial intelligence (AI). So artificial intelligence, machine learning, and deep learning can be represented by the following graph:

What is machine learning?

The basic idea of machine learning

1. Abstract real life problems into mathematical models, and understand the role of different parameters in the model
2. Using mathematical methods to solve the mathematical model, so as to solve the problems in real life
3. Evaluate whether this mathematical model really solves real life problems, and how well?

No matter what algorithm is used, what kind of data is used, the most fundamental idea can not escape the above 3 steps!

When we understand this basic idea, we can find:

Not all problems can be converted into mathematical problems. AI can’t solve real problems that can’t be transformed. And the hardest part is turning a real problem into a mathematical one.

Principles of machine learning

Take supervised learning as an example to explain the realization principle of machine learning.

Suppose we are teaching children to read (one, two, three). We will first take out 3 cards, and then let the children read the cards, while saying “a line is one, two lines is two, three lines is three”.

Repeat the above process and your child’s brain is constantly learning.

When repeated enough times, the child learns a new skill: one, two, three.

Let’s use the analogy of machine learning. Machine learning is similar to the human learning process mentioned above.

• The cards mentioned above are called training sets in machine learning

• The aforementioned “one horizontal line, two horizontal lines” attribute that distinguishes characters is called — feature

• The process of continuous learning is called modeling (Training model)

• The rules that emerge from learning to read are called models

Through the training set, constantly identifying features, constantly modeling, and finally forming effective models, this process is called “machine learning”!

Machine learning training methods

Machine learning can be roughly divided into three categories according to training methods:

Supervised learning

Unsupervised learning

Reinforcement learning

Supervised learning is when we give an algorithm a data set and give it the right answer. Machines use data to learn how to calculate the correct answer.

Here’s an example:

We had a whole bunch of pictures of cats and dogs, and we wanted the machine to learn how to recognize cats and dogs. When we use supervised learning, we need to tag the photos.

We tag photos with “correct answers,” and with a lot of learning, the machine can learn to recognize cats and dogs in new photos.

This way of helping machine learning with lots of manual tagging is called supervised learning. This way of learning works very well, but it is also very expensive.

In unsupervised learning, there is no “right answer” to a given data set; all data are the same. The task of unsupervised learning is to mine potential structures from a given data set.

Here’s an example:

We gave the machine a bunch of pictures of cats and dogs, and we didn’t label them, but we wanted the machine to sort them.

The machine was taught to categorize the photos into two categories, one for both cats and one for dogs. Although the results of supervised learning look similar to those above, there are essential differences:

In unsupervised learning, the machine did not know which was the cat and which was the dog, although the pictures were divided into cats and dogs. For machines, there are two types: A and B.

Reinforcement learning is closer to the nature of biological learning and therefore is expected to achieve higher intelligence. It focuses on how an agent can take a series of actions in the environment in order to achieve the maximum cumulative return. Through reinforcement learning, an agent should know what behavior to take in what state.

The most typical scenario is playing games.

On January 25, 2019, AlphaStar defeated professional starCraft players TLO and MANA.

7 Steps of machine learning practice

Machine learning can be divided into seven steps in practical operation:

1. To collect data
2. Data preparation
3. Choose a model
4. training
5. assessment
6. Parameter adjustment
7. Predict (come into use)

Let’s say our task is to distinguish wine from beer by alcohol and color. Here’s a closer look at how each step in machine learning works.

Step 1: Collect data

We buy a bunch of different kinds of beer and wine at the supermarket, and then we buy spectrometers to measure color and equipment to measure alcohol.

At this point, we mark all the wines we buy with their color and alcohol level to form the following table.

color

alcohol

species

610

5

beer

599

13

Red wine

693

14

Red wine

…

…

…

This step is important because the quantity and quality of the data directly determine the quality of the prediction model.

Step 2: Data preparation

In this example, our data is very neat, but in the actual situation, there will be many problems in the data we collect, so data cleaning and other work will be involved.

When there was no problem with the data, we divided the data into three parts: training set (60%), verification set (20%), and test set (20%) for subsequent verification and evaluation.

Step 3: Select a model

Researchers and data scientists have created many models over the years. Some are great for image data, some are great for sequences (like text or music), some are for digital data, and some are for text-based data.

In our example, since we only have 2 features, color and alcohol, we can use a small linear model, which is a fairly simple model.

Step 4: Training

Most people think this is the most important part, but it’s not — the quantity and quality of the data, and the choice of models, are more important than the training itself (3 minutes on the training platform, 10 years off the stage).

This process does not require human involvement, the machine can do it alone, the whole process is like doing arithmetic. Because the essence of machine learning is the process of turning problems into mathematical problems and then solving mathematical problems.

Step 5: Evaluate

Once the training is complete, the usefulness of the model can be evaluated. This is where our previously reserved validation sets and test sets come into play. The evaluation indexes mainly include accuracy rate, recall rate and F value.

This process allows us to see how the model makes predictions about numbers we haven’t seen yet. This is meant to represent how the model behaves in the real world.

Step 6: Adjust parameters

After completing your assessment, you may want to know if your training can be further improved in any way. We can do this by tweaking the parameters. When we train, we implicitly assume parameters that we can tweak to make the model perform better.

Step 7: Predict

The six steps above serve this purpose. This is also the value of machine learning. At this point, when we buy a new bottle of wine, just tell the machine its color and alcohol, it will tell you whether it is beer or wine.

Classic machine learning algorithms

• Convolutional Neural Network (CNN)

• Recurrent Neural Network (RNN)

• Linear regression

• Logistic regression

• Linear discriminant analysis

• The decision tree

• Naive Bayes

• K neighboring

• Learning vector quantization

• Support vector machine

• Random forests

• AdaBoost

• Gaussian mixture model

• Limit boltzmann machine

• K – means clustering

• Maximum expectation algorithm