Machine learning is all about being able to teach computers how to “grasp” new concepts, but it often requires hundreds of examples during hours of training – all in all, a pretty inefficient process. But that’s up for change thanks to a new piece of research published on Friday.
In an attempt of shortening the learning process and transform it in something more alike to human thinking, a team of researchers has developed a Bayesian Program Learning framework. It only needs a few examples in order to apply new concepts and bits of knowledge, and the goal is to then teach computers to detect and replicate handwritten characters based on only one example.
There’s a great difference between the standard pattern-recognition algorithms – which present concepts in the form of pixel configurations and collections of features – and the BPL approach. The latter basically “explains” the algorithm all the data that it needs to learn, such as the character.
After presenting the concepts as “probabilistic computer programs,” the algorithm has the ability of essentially programming itself by writing the code it needs in order to produce the letter it identifies. So far, the BPL can also reproduce the many variations people draw a given letter.
Learning new concepts is also a sped up process because the model “learns to learn” by recalling previously learned concepts. For example, it can use knowledge of the Latin alphabet as basis for learning new letters in the Greek alphabet a lot quicker.
Most important of all is the fact that the algorithm was able to teach computers how to pass a sort of “visual Turing test.” In other words, the researchers pitted humans against computers in a task of reproducing a series of handwritten characters – after just seeing one example of each.
In some examples, both human and computer subjects were tasked with creating new characters based on the style of those previously shown. Conclusions showed that human judges couldn’t tell the difference between the human and artificial intelligence results.
This model was tested on more than 1,600 types of handwritten characters covering 50 writing systems, such as Tibetan, Sanskrit, Gujarati and Glagolitic. Surprisingly, the model did rather well on reproducing invented characters (think in the range of Futurama).
Researchers responsible for creating the model said that “it has been very difficult to build machines that require as little data as humans when learning a new concept.” That’s why it’s so exciting that research has birthed a way to replicate these abilities.
Image Source: Square Space