In the third order of the series, we continue to look at the development history of artificial intelligence from percent theory to the present day. If you haven't checked out the past content of the series yet, I recommend reading it first.
[AI Story] Crucial Moments of Artificial Intelligence 1
[AI Story] Crucial Moments of Artificial Intelligence 2
2000s, the advent of deep learning
Artificial neural network research in the 1990s had to go through a difficult period due to limitations such as loss of tilt problem*. It was the second unusually cold and long winter of artificial intelligence that continued for more than 10 years. Once again, the world's interest and investment disappeared, and many researchers had to leave.
However, in the 2000s, a revival period came. It was thanks to a few pioneers who endured difficult times and continued their research. Also, with the spread of the Internet, big data was created, computer performance improved, and problems with existing algorithms were solved, setting the stage for a new leap forward.
Finally, in 2006, Jeffrey Hinton published a paper called 'A Fast Learning Algorithm for Deep Belief Nets' **. We solved the challenges of existing neural networks through a new algorithm called Deep Belief Network (DBN) ***. As a result, the era of deep learning, which continues to this day, will begin in earnest.
DBN, which built up RBM unsupervised learning methods and RBM, sources
A deep trust neural network consists of a restricted Boltzmann machine (RBM) **** composed of an input layer and a hidden layer, stacked on multiple layers like a building block (building block)*****This is it. We were able to solve the problem of loss of slope through pre-training (pretraining), and even solve the problem of not being able to process new data well by using a method (dropout) ****** that deliberately omits data during learning. Through this method, it was possible to create deep learning beyond the limitations of existing artificial neural networks.
What's interesting is that deep learning is actually another name for artificial neural networks. It is said that the image at the time was so bad that the paper containing the word artificial neural network was rejected by the title alone. ******* In other words, an unavoidable choice to escape negative perception has led to deep learning today.
2012, AlexNet wins ILSVRC
The developers of AlexNet - Ilya Sutskever (left), AlexKrizhevsky (middle), Geoffrey Hinton (right), sources
In 2012, AlexNet won ILSVRC (ImageNet Large Scale Visual Recognition Challenge) ********. ILSVRC was an image recognition contest that competed for the accuracy and speed of algorithms. AlexNet, created by Jeffrey Hinton and his students, won the championship with an overwhelming result that lowered the previous year's record by 10% or more.
AlexNet was very different from the previous system. Deep Neural Network (Deep Neural Network) was implemented using a spiral neural network (CNN), an artificial neural network model based on the structure of the human brain. ********* Deep Architecture (Deep Architecture) -based deep learning algorithms became mainstream after AlexNet.
By the way, I already read about AlexNet in a previous post. If you want to know more about AlexNet's innovation, which opened the heyday of deep learning, check it out as well.
[AI Story] AlexNet (AlexNet) opened the era of human vs. artificial intelligence (3) deep learning
2016, AlphaGo wins with the highest number of humans in Go
The scene of AlphaGo and the 9th Dan Lee Se-dol, sources
AlphaGo, which won the Go match with Lee Se-dol in 2016, is still synonymous with artificial intelligence. As such, AlphaGo left a strong impression on us at the time. In fact, AI technology began to be prominent in various fields after AlphaGo.
AlphaGo used a new learning method that combines traditional supervised learning and reinforcement learning. In particular, they improved the accuracy of predictions through self-publishing rather than just learning by looking at reports. Also, the optimal number was determined using Monte Carlo tree search techniques suitable for Go, where a large number of cases unfold. **********
AlphaGo's victory was a defining moment that will continue to be remembered as a highlight in the history of artificial intelligence. It caused tremendous repercussions not only in the field of artificial intelligence research, but throughout society. It reminded me that even intuition and reasoning, which until now are unique human abilities, can be calculated by AI.
I've already read about AlphaGo in my last post. If you have any questions about AlphaGo's algorithms and learning methods, check it out as well.
[AI Story] Humans vs. Artificial Intelligence (5) AlphaGo (AlphaGo) calculating intuition
While finishing
Artificial intelligence began with an attempt to simply implement neural networks in the human brain. In other words, the history of AI was an adventurous journey for machines to resemble human brains. However, the human brain was an unknown world that was too complicated to imitate, and AI reached today through many trial and error and several decisive moments.
In the future, it seems that the evolution of artificial intelligence will not just beat humans in Go. In fact, AI is already being used to solve various problems in reality. In addition to weather prediction, language understanding, medical research, etc., it has a wide range of applications, from automobiles to toys, etc., and is showing its presence.
Recently, artificial intelligence is developing at a very rapid pace. Right now, even experts say it's hard to predict a few years from now. It is a situation where expectations and fears about AI that will surpass humans, and mixed views on the future with AI intersect.
Demis Husavis, CEO of DeepMind, who developed AlphaGo, said, “In the future, artificial intelligence will help humanity pioneer new fields of direction and discover truth.” At the same time, I also said that it's just doing what humans have done, and that developing AI with imagination and creativity requires a deeper understanding of the human brain. *********** If so, why not put aside hasty predictions and carefully watch what other decisive moments AI will evolve in the future?
**** For Boltzmann machines and RBMs, please refer to the following article. 'Boltzmann Machine: Principles of Generative Models' https://horizon.kias.re.kr/18001/
References
[1] https://ko.wikipedia.org/wiki/인공지능#역사
[2] Introduction to deep learning and major issues https://www.koreascience.or.kr/article/JAKO201525257248863.pdf
[3] https://en.wikipedia.org/wiki/AlexNet
[4] Jeffrey Hinton, the godfather of artificial intelligence in the 21st century, professor at the University of Toronto, Canada https://www.joongang.co.kr/article/20382230#home
[5] https://ko.wikipedia.org/wiki/알파고
[6] From ImageNet in 2012 to AlphaGo... all about deep learning https://blogs.nvidia.co.kr/2016/03/21/all_of_deeplearning/
Good content to watch together
[AI Story] Crucial Moments of Artificial Intelligence 1[AI Story] Crucial Moments of Artificial Intelligence 2
