A child will just see an elephant once to identify all elephants will see in the future. Whether Africans or Asians, or see a movie at the zoo, in the case of a herd on the savannah or one behind the trees, you know you’re seeing elephants. So far, the artificial intelligence needed thousands of images of elephants in all situations to identify new one that was not in their database. Unlike humans, he was not able to generalize based on a few examples. However, this type of human learning as just be replicated by a machine.
A group of American researchers has created a mathematical algorithm that allows machines to learn how a child. This inductive way of acquiring new knowledge is one of the strengths of human versatility. Before a new concept or object, a few examples are sufficient, sometimes only one, to extract from them the basic elements of the object and the relationships between its parts. In this way, the child just need to distinguish an elephant training of a mammoth. More importantly, this type of learning is inside the seed of creativity. Knowing what is an elephant, humans can imagine new elephant examples, including a pink elephant flying.
“There are many learning systems [machines],” says the professor of the University of Toronto (Canada) and co-creator of the algorithm, Ruslan Salakhutdinov. “It usually takes hundreds of thousands of examples to train the concept one wants to learn. But humans are able to capture these similar categories, such similar concepts, with just a few examples, if not a unique training” Salakhutdinov added, considered one of the pioneers of artificial neural networks key to machine learning (or machine learning, by its original terminology in English).
as powerful as the Watson supercomputer systems, recognition Face Facebook or personal assistant Siri Apple use this machine learning based on hoarding as many examples as you can, collect a lot of data and relate them algorithmically.
The algorithm created by Salakhutdinov and two colleagues from US universities not intended to see elephants but a machine that can identify handwritten characters about 50 writing systems, from the Greek alphabet to Sanskrit, going through some invented, like the Futurama series. This is a library of 1,600 different character types. The potential is enormous variability and that regardless of possible fonts (Arial, Comic Sans, Helvetica …) in each system or the calligraphic style of writing.
“The idea for this came algorithm a surprising finding we made while recopilábamos a database of handwritten characters around the world, “explains the researcher from New York University and co-author Brenden Lake.” We saw that if you ask a group of people draw a new character, there is a persistent pattern in the way they do: they tend to create new characters in the same way, based on the parties or lines that have been drawn before, “he added.
The algorithm presented in an article in the journal Science operates in a similar manner. After showing a written character for one or two different hands, the machine breaks down into its fundamental parts and find the relationships between them. In this way, the system can identify dozens of new versions of a character and even new characters of the same alphabet.
To check the validity of their algorithm, the researchers asked a group of humans hand to write a character string. The machine was instructed to do the same. After a series of human judges familiar with the script used in each of the experiments, was trying to distinguish which characters had been written by a human and which by the machine. In most cases, they were not able.
Although the algorithm is designed for handwritten characters, the researchers believe their approach can be applied to other fields such as computer vision, recognition voice or natural language processing. As Salakhutdinov says: “We hope this work will help guide the advancement of artificial intelligence, developing a new generation of smart systems, smart machines that can deploy or less close to human intelligence”
. The algorithms do better every time
As noted by Professor Laboratory of Artificial Intelligence and Computer Science Massachusetts Institute of Technology, the Barcelona Antonio Torralba, “learn a few examples was one of the initial objectives artificial intelligence. ” So you think this new work as relevant. “They have built a system with the basic rules by which a new object is made. Once these basic rules are known, can learn new objects because you have little to identify,” says the Spanish expert in machine learning and computer vision.
In the case of the characters, there are traces, leading them to a broken model, and many of them are composed of similar elements, such as straight or circular lines. Once discovered these basics, before a new character just have to find out what elements compose. Torralba, “if you get to learn with a few examples, we can build machines that can work closer to the human being, able to learn and react with very little data, as humans do.”
The progress of AI is reducing action areas uniquely human. Overtaken long ago by the machines in computing power, capacity to store information or to establish relationships between data, humans are now becoming tuteados intelligence.
Humans do worse in the stock market, where at least in the United States, more than 75% of transactions the machines make. Lending, operations in the operating room or harvesting plant are increasingly automated decisions. The algorithms are also behind the recommendations of reading that Amazon, the pairings in the pages of online dating or driving vehicles without driver (human) that will begin soon to reach the road.
“Humans should drive cars? I think not and that should be banned,” said the expert in big data and machine learning of ASPgems, Juantomás Garcia. For him, the algorithms are not reducing the sphere of human action, but that is improving. As he says, “there are millions of shares does not make sense that we continue making human”.
Miguel Angel Criado / El País
No comments:
Post a Comment