MatSQ Blog > Column

[MATSQ COLUMN] #002. How Far Can Artificial Intelligence Technology Go?

2021-01-08 08:46:02

What are the most striking developments in recent years?

The first that came to mind is the visible accomplishments of artificial intelligence (AI) featuring AlphaGo. The reinforcement learning of artificial neural networks currently used is considered the highlight of inductive reasoning developed by humankind. It is at least true if only taking into account deep learning that has been recognized since 2012. As is well known, AI has gotten to where it is today through two AI winters or, more appropriately, two dark periods. Until 2012, it had been through a brutally long maturing time. As some simple but innovative ideas emerged around 2012, it was found that well-known artificial neural networks could perform much better, which we call deep learning.

 

The advent of AlphaGo in 2016 came as a fresh shock as it was introduced to Koreans through a grand event. Moreover, it seemed to be appropriate for the general public. On the other hand, stubborn experts revealed their skepticism about AI. Such people had already succeeded in a frame based on deductive reasoning and could live well without AI. For them, opening a new marketplace was little more than adding one more irksome work. However, the same did not go for all people, regardless of time and place. Deep learning looked like such a comprehensive approach because, in so many fields, prompt applications of the technology were as natural as breathing. Still, some experts in specific fields were facing many practical difficulties in adopting it for their problems.

“To be honest, I believe AI was able to succeed in the special situation of the game of Go! Wasn’t it?”

These assessments seemingly still hold true so that many, including myself, have not been able to show hostility. Perhaps there was “insufficient evidence” achievements then.

 

Then, time passed again. A year later, in 2017, AlphaGo Zero burst onto the scene, which was even more shocking. The significance of AlphaGo Zero’s appearance has been largely overlooked by both the general public and experts. However, it forced people to erase something like the last unwritten rule that artificial neural networks must learn known data from their minds. Prejudice is eradicated through innovation. That is, AI was no longer applied only to simple hyper inductive thinking that can be established confined to existing data. Through AlphaGo Zero, we witnessed that the “God of Go” could be born just by applying the basic rules of Go. “Why on earth is human data required to learn the game of Go?” In fact, it could be considered the theoretical conquest of Go, with its long history of 5,000 years. As of 2020, a decent AI Go program has never lost to any professional player.

 

 

In 2017, the AlphaGo Zero development team had emphasized that the game of Go could be conquered with the so-called first principles.

In other words, it starts from scratch by training itself without any data from human games. Because AlphaGo Zero is a self-learning system, it has “Zero” in its name and uses the first principles strategy. The team stressed in a paper that AlphaGo Zero starts from "tabula rasa", without a set opinion. It means it can learn the principles through sufficient self-studying. Still, two more things need attention: the speed and level of learning. Self-teaching is not only faster than learning from human data but also allows a higher level of learning. AlphaGo Zero discovered formulas that humankind has never found. Therefore, AlphaGo Zero could demonstrate a new level of Go ability. The technology is no longer constrained by the limits of human knowledge, which it transcends through its nonuse of data from human games.

“How could this be possible? Without watching the plays of Pele, Cruijff, Maradona, Ronaldinho, Cristiano Ronaldo, and Messi, is it possible that a greater football player is born?”

To some extent, I believe an old idiom that “no matter how difficult it is, one-hundred-time readings enable you to understand it.” , but not fully. As far as I know, it can be applied to get already unknown formulas only. As such, you cannot answer many questions without application skills in certain problem solving, even more so if you don’t have a high degree of concentration and insight to calmly count a tangled skein. Professors say that students cannot solve a problem at all if there is a small variation. As is well known, a lack of high level of learning leads to a lack of creativity, with which you cannot find anything new. Sometimes, therefore, a lack of creativity causes the greatest pain.

 

 

In the meantime, even more amazing things were underway.

Shortly after AlphaGo Zero came out in 2017, Demis Hassabis from DeepMind shared his intention of solving protein folding and material design. Then, I said to myself, “Right!” As a total stranger to the game of Go, I thought that at least such a level of problems would be worthy and fun to solve. Time passed again, and DeepMind brought out a new algorithm called AlphaFold in 2018. It breathed new life into the field of protein structure prediction. In a sense, it offered a possibility of solving difficult problems. Exactly two years later, in 2020, DeepMind finally made a huge positive impact on experts who had studied predicting protein structure with AlphaFold 2. It predicted protein structure, which was classified as the most difficult problem, to a satisfactory level. After about 50 years of trials, Google (an affiliate of DeepMind) finally succeeded in developing a method for predicting the structure of a folded three-dimensional protein from amino acid sequence information at an acceptable level. This was a remarkable scientific achievement that universities and research institutes had failed to accomplish. It also helped humankind to take one step toward the fundamental understanding of life phenomena, as proteins can perform exactly the “programmed” life phenomena without an error only if they have their own three-dimensional structure. This scientific achievement can be applied immediately to overcome diseases and develop new drugs.

In all ways, Google’s scientific accomplishments from AlphaGo to AlphaGo Zero, AlphaFold, and AlphaFold 2 are truly amazing. The academic successes achieved by a company rather than universities and research centers should never be overlooked. One must note that even the most prominent experts have not achieved such.

 

Furthermore, AI has a sufficient reason to exist only with the ability to do better than humans in a specific field. AI never feels tired and can produce outcomes without a break. However, this is an underestimation of AI’s purpose and capabilities. It can serve as a powerful foundation required for more creative work and can make technological breakthroughs necessary for specific fields. As mentioned above, AI has demonstrated its ability to figure out things that humankind has not solved for 5000 and 50 years, respectively. As of 2020, there are three perspectives on AI. First, “It has nothing to do with me.” Second, “I have tried using it, but it doesn’t mean much to me.” Third, “it would be beneficial if I can put it to good use.” Which category do you fall under? Again, time is passing by.

How far can AI technology go? Who could ever answer this question confidently?

Perhaps people with the third perspective may answer.

 

 

 

Author

In-ho Lee

Principal Researcher, Korea Research Institute of Standards and Science

 


Comments