It is extremely difficult to make machines learn about the world around them. The learning processes for artificial intelligence (AI) often involves a huge amount of data and ample of learning time. But with a new algorithm that has been developed by scientists in the United States, the time it takes for an AI to be taught things can be cut down to a great extent. This algorithm will assist machines in being capable of learning things the way humans do by recognizing and drawing visual symbols. Also, the symbols drawn by the machines will be nearly impossible to differentiate from those drawn by human hands.
Their research says that humans are good at learning things because of their imperfections be it about learning a character, the way to operate some tool, some dance move or anything. Humans just need to see and analyze a few examples and they can very easily replicate what they have been shown or taught.
Whereas machines take time to learn about every single curve when being taught to identify faces or any written or typed characters. Machine learning involves hundreds or thousands of examples before the AI finally becomes accurate.
What differentiates human learning from machine learning is that humans just observe how a character or a figure needs to look like and they draw it their own way, which never looks exactly the same way as the character that has been shown to them. But machines work on creating an exact replica of the image shown to them.
But with this new probabilistic algorithm developed by the researchers in the US, rather than reproducing exact replicas of the character learned every single time, the machines would now be able to recognize the curve and draw it slightly differently in every instance, much like humans. Researchers have come to develop this algorithm using Bayesian program learning framework.
This algorithm was then exposed to 1,600 types of handwritten characters taken from 50 writing systems from the ones being used in the world which included characters from Sanskrit and Tibetan. The machine was made to recognize the characters and learn to draw them independently.
After this, researchers conducted ‘visual Turing test’ where human judges had to differentiate been machine and human drawn characters. But even less than 25% judges performed a bit more than pure by chance guess work in distinguishing the drawings apart. All results from the test and findings have been reported in ‘Science’.
Joshua Tenenbaum, a researcher in cognitive sciences at the Massachusetts Institute of Technology (MIT) said a single example is enough to teach human children new things even before they go to prep school. They are also capable of imagining new examples which they haven’t even been shown. According to Joshua, we are still far from developing machines which could learn as smartly and as fast as a human child does.
But with this algorithm a machine was able to learn through real world concepts for the first time ever. Researchers are working to improve a machine’s ability to learn by using such fast study algorithms. This would be a huge success on the artificial intelligence front if machines could grasp new concepts faster such as processing images, recognizing speech, faces, understand language naturally and retrieve information from the real world.
Article By: Dakshita