Artificial Intelligence Learns From its Mistakes

It is well known that we all learn from our mistakes. All creatures from the animal kingdom answer a causality principle which forces them, by memory, reason or instinct, not to reproduce an error if they have already been subjected to a physical or psychological reprimand. Since this type of perception is specific to beings endowed with feelings and consciousness, it seems impossible to apply it to computers. But what about artificial intelligence ?

more

Nowadays, AI is not only limited to the recognition and processing of programs made to achieve a specific goal, one action at a time. In 1951, at the beginning of AI, a student from the University of Manchester created a machine capable of beating you hands down at the game of chess. This is an example among many others of primitive artificial intelligence, outstanding at the time, but quickly outdated by technical and scientific needs.

Through the years, scientists have discovered the need to revolutionize AI and decided to use the most logical example to leverage their advances : the human being. The tendency to make mistakes is thus an imperfection envied by machines. Current artificial intelligence systems are therefore capable of reproducing thoughts, a deductive mind and even reasoning in order to stock each piece of data into its own memory : an advanced learning process.

In the past machine translation software used to run with an algorithm splitting the whole text into fragments and then searching the meaning of these fragments in the memory. Depending on the language, the software would then adapt the structure of the fragments based on the grammatical rules of each language. But since the translation’s quality was not always perfect, online translation has recently been suited with a new advanced learning system, allowing the software to learn from its mistakes.

As an example, let’s take two languages intrinsically different : English and Japanese. Having encountered performance difficulties during a previous translation in this language combination, the software changes its strategy and chooses a “compromise” language. In this case, Korean acts as a bridge between two grammars that are too different. After a short analysis, the automated translation software first translates from Japanese to Korean, and then from Korean to English. It gets around the difficulty and improves the result.

By taking language as an example, we quickly understand how much modern AI learning, which imitates human reasoning, can become effective. We could think that the gap between AI and human beings is becoming smaller, but the non-mechanical mechanism of the human mind cannot be excluded from the equation. The nuances, the feelings, the cultural knowledge, etc. What makes humanity beautiful is the myriad of unsolvable enigmas that even the most powerful calculator,-let’s call it “computer”- could not solve.

Written by Gildas Mergny

Translated by Arthur Chevallier-Letort

Leave a Reply