MENU

AI Reconstructs Periodic Table In A Few Hours, Humans Took A Century

Close

An artificial intelligence program has recreated the periodic table of elements in a few hours, a feat that humanity took almost a hundred years to accomplish.

Stanford University researchers have developed a new AI that was able to organize the entire periodic table without supervision simply by analyzing a list of chemical compounds.

Study leader and physics professor Shoucheng Zhang of Stanford's School of Humanities and Sciences hope the AI could go on to set the new standard in machine intelligence.

Borrowing Concepts From Natural Language Processing

By borrowing concepts from natural language processing, which helps machines analyze, understand, and use human language, the AI called Atom2Vec was able to reconstruct the periodic table from scratch.

More specifically, the program took advantage of the idea that the meaning of a word can be deciphered simply by analyzing the other words around it. In a matter of a few hours, the AI was able to arrange all 118 elements in the periodic table according to their several properties. In contrast, humans took nearly a century to organize the elements.

The results of the new research are published in the Proceedings of the National Academy of Sciences.

Proposing A New Gold Standard Of Machine Intelligence

The ultimate goal for Atom2Vec is to establish a new benchmark of intelligence for AI. The current standard for machine intelligence is passing the Turing test, which was devised by computer scientist Alan Turing in 1951.

Essentially, the Turing test consists of a human sitting behind a computer screen who serves as the judge. The judge chats with a group of anonymous humans and a chatbot, designed to trick the judge into thinking that it is human. If the chatbot succeeds, it passes the Turing test.

Although this method has been passed down through the decades, it has its obvious flaws. Humans, first of all, are irrational, which means an AI has to develop these human irrationalities to pass itself off as a human.

"That's very difficult to do," says Zhang, "and not a particularly good use of programmers' time."

In the place of the Turing test, Zhang wants to advance a new way to test machine intelligence. For AI to be smarter than humans, they have to be able to discover a new law of nature. However, before a machine can achieve this lofty accomplishment, Zhang wanted to see first if it can replicate one of man's greatest discoveries. Apparently, Atom2Vec passed the secondary test.

How AI Organized The Elements

The team programed Atom2Vec after Word2Vec, an AI program developed by Google engineers which turns words into codes called vectors. By studying these codes, the program can predict the meaning of a word based on the other words that appear close to it.

For example, the word "king" often appears with "queen" in the same way as "man" is associated with "woman". Therefore, the code for "king" would be something like "king = queen minus woman plus man".

Borrowing this concept, the researchers programmed the AI to analyze a list of chemical compounds to see if it can decipher the elements and their various properties.

For instance, looking at a list that contains sodium chloride (NaCL) and potassium chloride (KCl), the AI can deduce that sodium and potassium belong to the same group because they can both bind with chlorine.

Future Use For Atom2Vec

The researchers have already thought up a potential application for the new AI and are, in fact, already working on it.

Zhang says Atom2Vec can be programed to find the right antibody that can fight off cancer cells. One of the most promising treatments for late-stage cancer is immunotherapy, which uses the patient's own antibodies to attack tumorous cells.

The challenge is in finding which of the 10 million unique antibodies are right for the job. The researchers say that their AI would be able to classify all 10 million antibodies and organize them into something like the periodic table of elements.

If doctors find a new antibody that is effective but may be harmful to the patient, they could easily look at antibodies with the same properties for a new solution.

ⓒ 2018 TECHTIMES.com All rights reserved. Do not reproduce without permission.
Real Time Analytics