반응형

Until last time, I briefly learned about the history of artificial intelligence and artificial intelligence. Now let's find out what machine learning is in earnest.

Machine Learning

= Self-learning technology to improve performance

Let's first pick up the relationship between artificial intelligence and machine learning. Machine learning is a field of artificial intelligence, a technology that allows machines to learn and perform better on their own. This technology is a small category of artificial intelligence.

Machine learning is called machine learning in Korean. What is this machine learning? To put it simply, the program gives input to the program, checks the output, adjusts the logic according to the output, and improves itself by modifying the logic to produce better results when the program gives the next input.

Self-correcting logic?

Non-Communists and those who have done some coding can be a little confused meaning up there. You're modifying logic on your own? The traditional program we know is that humans plan logic and print out 1 if the input value a is greater than 10! Otherwise, print zero!!

input a 
if( a > 10 ){ 
  print : 1 
}else{ 
  print : 0
}

If you squeeze and spin logic like this, it's only by that logic. By the way, this logic is being modified by a computer? That's right. If you say something, the program itself doesn't change the source itself. The method is to save the figures through mathematical modeling and update them to find the best value.

For example, I entered 3 in the input, but the result is 0? I entered 11 in the second input, but the result is 1? I entered 10 in the third input, but the result is 0? Based on the results of the three inputs, we found that if the input was greater than 10, we would return 1 and if it was smaller, we would return 0. Then the program sets the standard as input>10. The program modifies this reference value by itself through the mathematical model mentioned above, even if the output and other inputs later.

To sum up, if the traditional program of the past was for the user to enter the rule and produce the output, machine learning is for the user to enter the output and make the rule with his own thread. Do you understand a little bit now?!

There are many kinds of machine learning. Logistic regression, linear regression, artificial neural network, etc. Let's just go over the learning types of machine learning.

Meaning of training in machine learning

In machine learning, the word training is used a lot. No, it's no exaggeration to say it's a necessity. So what does the word training mean in machine learning? It's exactly what we know and what training means. If you take a quick look at how machine learning works, it gives input and the program checks the input and output and modifies the rules on its own. This one-time course is one-time training.

If it gives 100 inputs and outputs and the rule changes 100 times, it is said to have trained 100 times. In other words, training is the process of giving various inputs (problems) and outputs (answers) and modifying standards so that the program can improve itself. Literally, machine learning is a machine learning process, so training is essential in this learning process.

instructional learning
= supervised leading
= train models with input and target

 

Map learning is literally human guiding programs if you understand them. In the example of the above data, the program gives past data on whether it actually rained, depending on the amount of clouds, which is called training data. We put in a total of 7 data, 67% of the cloud was not enough, 31% was not enough, 99% was raining. The program trains a total of seven data sequentially.


Eventually, after seven training sessions, Plgram creates a standard of rain if the amount of clouds is approximately 70%, or less. There are only seven data, but when hundreds of millions of training data come in, this standard becomes very sophisticated.

It is guidance that gives data on past inputs and outputs, and identifies patterns in this data to create a baseline. And if you put a new input into this trained data, it's the map learning that shows the expected results!

→Limits of map learning

The most important thing in guidance is training data. It takes a huge amount of training data to become a sophisticated program, and the process of collecting or creating such data is not easy, and it is also problematic when there is data with large errors.

The biggest problem is that the model is not updated in real time when new inputs come in. New training should include newly entered data, which means that once a model is made, it will only return the expected value for that model and it will be difficult to adapt to environmental changes.

non-map learning
= unsupervised leading
= Training with targetless data

As opposed to map learning, there is no target data. In other words, there is only an input and no result data. There is a group classification as an example of using this Bizzy learning.

It is like no target because no group exists until a group is created, and there may be any group. Without a target, the program creates a cluster by classifying the inputs accordingly.

This is the way to learn Bizzy.

reinforcement learning
= reforestation leading
= reward and punish the result.

If you look at the instruction I saw earlier, they used inputs to create rules. However, when the nature of the input changed, there was a limit to not being able to adapt. However, using Ganghwa learning can solve the problem above to some extent.

Reinforcement studies evaluate targets with input. So if it is right, it rewards it, and if it is wrong, it punishes it and adjusts the rules it has. This allows the rules to adapt to the environment by reward and punishment even if the input's personality changes later.

Making the above contents into a single model is called an agent. The agent's goal is to be rewarded as much as possible. Google's AlphaGo was also created using such reinforcement learning.

[Deep Learning] #3 Let's find out what machine learning is! / Relationship with Artificial Intelligence / Map Learning / Non-Map Learning / Reinforcement Learning

 

#DeepLearning #MachineLearning #ArtificialIntelligence #Relationship #Learningforguidance #Learningfornontraditionallearning #Studyingforstrength

반응형
반응형

We studied what artificial intelligence is last time. Simply put, what artificial intelligence was, you can sort it out to a "system for creating human intelligence."

The history of artificial intelligence began in the 1950s when Turing machines came out. Later in 1965, a chatbot system called Alisa was developed. Alisa is also known as artificial intelligence, but it was not the kind of AI that we thought would acquire knowledge by learning for ourselves because it was actually just a program that could branch out every situation and answer any situation.

XOR problem of Perceptron

= Cannot process beta logical circuit

As the history of AI begins, the development of AI, which seemed to be successful, faces a problem and faces a recession. On the one hand, it is said to be an XOR problem of Perceptron, and those who first started artificial intelligence now will naturally not know what Perceptron is, so you can just ignore it. I'll post it later, but right now, you can think of it as a solution that mimics a tiny neuron to implement AI.



Looking at the graph of the left-most XOR gate, there was a problem in which the results of + and - could not be divided into one straight line (linear). If you look at the OR gate and the AND gate can distinguish between + and - with a single line, XOR is impossible!! So I wondered if the history of AI, which had been popular, would end like this, but Professor Marvin Minsky said, "We can solve it using multi-layer."

MLP (Multilayer Layer Perceptron)

Using two layers (blue) instead of one-layer layers as shown above, it has been revealed that XOR results can be divided into two straight lines, as in the far right graph. However, at that time it was theoretically possible to model the above concept, but it was too complex to actually implement it. (You don't have to understand what the picture above is. I will study properly later on, so non-specialists just have to be aware that there is a model in shape.



In addition, he pointed out the limitations of this MLP model because it is a model that cannot be taught. However, this problem is solved through the reverse error propagation method in the future.

AI stagnation period

1970~1990

That's how the first AI winters, or AI winters, came in 1974 to 1980, and then the second AI winters from 1987 to 1993. This is the time when AI's development was stagnant.

 

In a chess game called DeepBlue, which was created by IBM in 1997, he defeated then-world chess champion Gary Kasparov in a showdown in a game he thought was only human territory. But at the time, Deep Blue was able to win by learning all the cases that could come out of the chess game, but if you put them into Go, the number of cases would be as large as the number of atoms in the universe, making them impossible to learn.



Thus, models that investigate and learn all situations like Deep Blue have limitations in applying them to other areas, which was shocking to the public, but not enough to change the paradigm academically.

the advent of the Neural Network

Artifical Neural Network : ANN

Over time, a model called artificial neural network came out. Using this model, we were able to solve the XOR problem in the way Professor Marvin Minsky said, using multi-layers. (Let's find out more about Neural Networks in future postings.)

In 2009, Google started the car in the above manner to build a self-driving car, and time passed so that AlphaGo of DeepMind, a subsidiary of Google, won the Go field that it thought was a real human domain in 2016, shocking the world.

Since then, there is no area where AI can't do it anymore. It has been confirmed that A can also be used in the realm of creation, which was thought to be a human domain. So far, we have briefly learned the history of artificial intelligence. Let's talk about machine learning and deep learning in earnest next time.

[Deep Learning] #2 History of Artificial Intelligence / XOR Problem of Perceptron / Artificial Neural Network (ANN)

#Deeplearning #Basic #Artificialintelligence #Perceptron #XORProblem #Artificialneuralnetwork #ANN

반응형
반응형

Recently, I have been interested in artificial intelligence, so I am studying it personally. We've seen artificial intelligence and words on the news, machine learning, and deep learning. Is this different? Just as we look at the forest to know one thing, let's first sort out the relationship between these concepts and start!

Artificial intelligence
=System for creating human intelligence

Artificial intelligence is also known as AI, which is interpreted as a "system for creating human intelligence

All of the AI Go players we know, AlphaGo and Iron Man's Jarvis, are AI. Yes, all human-generated virtual intelligence falls into a very large category of artificial intelligence.

Strong artificial intelligence vs. weak artificial intelligence

 

Strong Artificial Intelligence - Iron Man JARVIS

Strong Artificial Intelligence refers to artificial intelligence with strong intelligence, as called Strong AI, and AI with indistinguishable intelligence from humans. Iron Man's Jarvis is the most familiar river artificial intelligence to us. There is also a flexible conversation with Tony Stark, as if he were talking to a real person.

On the other hand, if you look at Samsung's Bixby and Apple's Siri, which we know as artificial intelligence secretaries, Iron Man's Java-like and flexible conversations have yet to take place with technology. Once a command is given by voice, it takes a second or two to recognize it, and the user always waits for the result of the command and says, "Well...Do you understand...? LOL. And set the alarm, call me, call me, or, to some extent, when I ask a whole new order or question, I say, 'I don't know yet,' or I give a completely wrong answer.

Yes, there is no strong artificial intelligence yet in 2020 when we live, and it requires and is difficult to implement that much. So nobody knows how to implement this technology, and it's hard to predict when it will be created.

Weak AI - Siri, Bixby, Tesla self-driving

 

As I said above. It is not yet equivalent to human intelligence, but simply assisting humans in judging or processing simple commands is all called weak artificial intelligence. Yes, this era of weak artificial intelligence has opened in 2020, and various kinds of artificial intelligence have been introduced, including artificial intelligence speakers, artificial intelligence secretaries Siri, Bixby, deepfake, imaging technology, artificial intelligence doctors Watson, and Tesla's autonomous driving.

The above features of artificial intelligence show that work is done in certain areas. The medical institution carries out relatively simple functions such as driving assistance, answering questions, and searching. Like this, we live in a world where weak artificial intelligence is to grow in earnest, and as mentioned above, strong artificial intelligence is a technology that is hard to think of now, so when ordinary people say AI, we can think of medicine artificial intelligence.

[Deep Learning] #1 Let's find out what artificial intelligence is / Strong artificial intelligence vs. Weak artificial intelligence

#Deeplearning #Introduction #Artificiaintelligence #strongartificialintelligence #weakartificialintelligence #AI

반응형

+ Recent posts