Being taught how to learn in the same way that we do, robots are becoming more and more humane. Scientists have developed a new algorithm which gives robots the ability of drawing and reading visual concepts almost in the same way as humans do.
Actually the algorithm has been developed by reverse engineering the way in which people are thinking when they try to solve a problem.
The new study has been published in the journal Science. The research represents the joint effort of a team of North-American scientists: Brenden Lake from New York University, Ruslan Salakhutdinov from the University of Toronto and Joshua Tenenbaum, from MIT.
Brenden Lake, the leading author of the research argues that the new method will also be useful to improve the learning abilities of other machines. The algorithm helps computers learn new concepts a lot faster than they did before and it also gives them the ability of being more creative.
According to the developers, average humans learn new concepts, such as a dance move or the working of a piece of simple equipment, after being exposed to them for only a few times. Until now, machines needed thousands of examples before being able to reproduce a similar task.
Salakhutdinov has published another paper about 10 years ago in which his research team has developed an algorithm which taught a computer to learn the handwritten digits from 0 to 9. The algorithm used a total of 60,000 examples.
The recent research aimed to shorten this time by making the algorithm more similar to human’s thinking. The scientists wanted to make the computer learn from a small number of examples and generate its own new examples for each concept or even entirely new concepts.
In order to achieve that, scientists have developed a new framework, called the ‘Bayesian Program Learning’. What BPL does is turning concepts into computer programs. New concepts are being coded by the computer without the presence of a programmer being needed. More than that, instead of simply reproducing the same concept, the computer generates new coding each time it runs, producing a different outcome. By doing this, it becomes capable of recognizing different variables of a concept. For example, reading different handwriting.
The revolutionary algorithm also learns itself how to learn so that the next time it comes over a concept similar to a previous one, it will use the previous model to learn the new one. For example, it can learn the Greek alphabet after knowing the Latin letters.
Image source: pixabay