With the Cognitive Computing Challenge well under way, contestants may have run into a common problem: machine learning algorithms require mountains of data in order to learn. Fortunately, a collaborative effort between researchers from the United States and Canada have developed software that learns more efficiently than ever before, meaning that it learns faster and requires less data than other algorithms. Developer Brenden Lake of New York University, along with Ruslan Salakhutdinov from the University of Toronto and Joshua Tenenbaum of MIT, recently revealed that their program can recognize and learn handwritten alphabetical characters. Since the programmers for the Cognitive Computing Challenge are creating programs that can scan and read documents, it's easy to see how handwriting recognition software would be an extremely powerful and useful addition in the field.
As reported by the MIT Technology Review, the new technique uses a Bayesian program learning framework that automatically creates a new program for each specific character. Then, it can run that program when it recognizes the same character in the future. That “recognition” is achieved with a probabilistic model that determines the likelihood that a particular letter is the same as one that has been seen before. This contrasts to previous character recognition programs that remembered specific pixels and matched them to unknown characters in order to identify them.
The researchers used the brain’s own learning capabilities as a model for their program, specifically the way that adults learn to read new languages. Tenenbaum explains that the program’s goal is not to recognize patterns like many other machine learning algorithms, but to “describe… the causal processes in the world…we’re trying to learn a program that generates those characters”. In other words, the program will be able to write as well as read the same way a human would.
The best programs for handwriting-recognition available today need to view hundreds if not thousands of characters in order to make even basic distinctions between letters. Any machine learning program that requires a ton of data is known as “deep learning”, and while this method is also based on organic cognitive capabilities, humans have additional, more sophisticated learning processes as well.
To test the program, the researchers had participants view several different characters, some written by humans and some by machines, and then asked them to determine whether the character was created by a human or a machine. Less than 25% of the participants were able to tell the difference. A key developer in the deep learning method, University of Toronto psychology professor Geoffrey Hinton is excited by the promise of this new research and believes that methods like this are compatible with deep learning and, when combined, could take advantage of the best that both have to offer.
Even skeptics like Gary Marcus, a cognitive scientist at New York University, see the program as a step in the right direction. While he doesn’t agree with the researchers’ concept of the way the brain learns, he thinks that there needs to be a way to solve machine learning challenges when there is very little data available. “This is proof you can learn faster. And I think that’s something people are going to think about”.
Marcus thinks that once machines can understand language, then we will have reached a paradigm shift in machine learning. For that to happen, however, they need to truly understand the meaning of language, not simply translate or reproduce it.
HeroX’s Cognitive Computing Challenge is the perfect opportunity to be an integral part of this paradigm
shift. Our contestant’s work has the potential to completely revolutionize any aspect of life that requires paperwork, so it is easy to see how ubiquitous it will be in the future.
The submission deadline is January 11, 2016, so make sure your work is ready to revolutionize the world! No pressure or anything.