Building machines with the intellectual capabilities of a human being is still underway, but a team of scientists from the University of Tortonto (Canada), Moore-Sloan Data Science at New York University (USA) and MIT have managed to emulate the ability of people to acquire new concepts from a single example thanks to a very particular algorithm. The study has been published in the journal Science .
The algorithm allows the system to self-program and use previously acquired knowledge to recognize and extract simple visual concepts just as we humans do. And is that for a computer, this possibility requires hundreds or perhaps thousands of continuous instructions to reach the precision that humans acquire so easily. The study “shortens the computational process of learning new concepts and broadens the application of machines to more creative tasks,” explains Brend Lake, leader of the work.
To achieve this technological milestone for the first time in history, researchers created Bayesian Program Learning (BLP) software, in which concepts are represented as basic computer programs. Unlike standard pattern recognition programs, this algorithm learns “generative models” of real processes, increasing its efficiency through better use of data.
In the same way, this algorithm learns to learn , for example, using knowledge of the Latin alphabet to learn the letters of the Greek alphabet. To test it, the software was applied to more than 1,600 types of handwritten characters with multiple writing systems, including invented or non-existent alphabets, as well as tests with human participants using tests similar to the well-known Turing Test.
The results of the Turing-type visual tests revealed that the BPL model achieved the same level of performance as humans and far exceeded any deep machine learning results obtained to date .
“Before entering kindergarten, children already know how to recognize new concepts from a single example, and they can imagine new examples that they have never seen. We are still a long way from building machines as intelligent as a child, but this is the first time that we have made a computer capable of learning and using a large number of real-world concepts, including simple visual concepts such as handwritten characters ”, explains Joshua Tenenbaum, a co-author of the study.
We are barely scratching the surface, but a day will come when machines will be able to learn from their environment just as people do.
Image Credit: Danqing Wang
