A team of researchers from the University of Toronto and the Massachusetts Institute of Technology (MIT) just got to create a computer model that is capable of emulating human capacity, so far only, to learn new concepts from a single example. The work was recently published in the journal Science.
Although the model can only, for now, learn the alphabet from handwritten characters, the underlying system could be expanded to build applications able to learn from others based on symbols, gestures, movements or systems languages, both written and Spoken. Quite a technological feat that marks the beginning of an era in which machines will be able to learn from their environment as do people.
New concepts
In recent years there have been significant and steady progress in methods of machine learning, but humans are still much better than when they assimilate new concepts. People, in fact, even children, need just one or two examples for this, whereas normally algorithms machines require dozens, or even hundreds, of different examples to "understand" a concept that had never faced before . What's more, once a person learns a concept for the first time, it is immediately able to apply it in many ways and different situations. Something that, until now, no machine was not able to do.
Indeed, it is precisely these uniquely human capacities which Brenden Lake and his colleagues have tried to "capture" in its new computer model. To do this, the researchers focused on a wide range of very simple visual symbols, such as handwritten alphabets worldwide characters. And they built the model so that I could learn this kind of visual signs, and carry out generalizations about them from a very small number of concrete examples. Scientists named their model BPL (Bayesian Learning Program), claiming that the program is able to "learn a whole category of visual symbols from a single example and then generalize from it in a way that is indistinguishable from that people use. "
After developing the BPL, Lake and his team approach compared how humans, BPL and other computerized systems faced learning five different challenges, including generating new examples of letters that have only been displayed a few times.
Like a person
Scientists tell how, in a challenging task of classification "at a glance" the BPL model got the same level of performance that humans and far exceeded any result of deep machine learning achieved so far.
In fact, the computer model was able to classify, analyze and recreate handwritten characters and even managed to generate new alphabet letters "correct" appearance as the test of Touring type with which the results of BPL were compared with those managed human test subjects. "In many cases He stated in the article Science- creative skills and generalization were indistinguishable from those that are characteristic of human behavior."
