Recently, a company has developed technology that allows a machine to learn effectively from small numbers of examples and hone its knowledge as more examples become available. It can be applied anywhere, such as teaching a smartphone to recognize user preferences or helping autonomous motor systems quickly identify obstacles.
The old adage “repetition is the mother of learning” applies perfectly to machines. Many modern artificial intelligence systems working in devices rely on repetition in the learning process. Deep learning algorithms enable AI devices to extract knowledge from datasets and then apply what they have learned to specific situations. For example, if you feed an AI system that the sky is usually blue, it will later recognize the sky among the images.
Complex work can be done using this method, but it certainly leaves much to be desired. But could you get the same results if you run the AI deep learning system through fewer examples? Boston-based startup Gamalon has developed new technology to try to answer this question, and this week unveiled two products that take a new approach.
Gamalon uses Bayesian programming techniques, software synthesis. It is based on 18th century mathematics developed by mathematician Thomas Bayes. Bayesian probability is used to make refined predictions about the world using experience. This form of probabilistic programming - where the code uses probable rather than specific values - requires fewer examples to infer, for example, that the sky is blue with patches of white clouds. The program also refines its knowledge as you further explore the examples, and its code can be rewritten to tweak the probabilities.
Probabilistic programming
While this new approach to programming still has challenges to solve, it has significant potential to automate the development of machine learning algorithms. “Probabilistic programming will make machine learning easier for researchers and practitioners,” explains Brendan Lake, a New York University researcher who worked on probabilistic programming techniques in 2015. "He has the ability to take care of the complex parts of programming on his own."
CEO CEO and Co-Founder Ben Vigoda showed MIT Technology Review a demo drawing application that uses their new method. It is similar to what Google released last year in that it predicts what a person is trying to draw. We wrote about it in more detail. But unlike Google's version, which relies on sketches already seen, Gamalon relies on probabilistic programming to try to identify key features of an object. Thus, even if you draw a shape that is different from the ones in the application's database, as long as it can identify specific features - for example, a square with a triangle at the top (a house) - it will make correct predictions.
Promotional video:
The two products presented by Gamalon show that their methods may find commercial applications in the near future. Gamalon Structure's product uses Bayesian software synthesis to recognize concepts from plain text and is already outperforming other programs in terms of efficiency. For example, having received a description of a TV from a manufacturer, she can determine its brand, product name, screen resolution, size, and other features. Another app - Gamalon Match - distributes products and prices in store inventory. In both cases, the system quickly learns to recognize variations in acronyms or abbreviations.
Vigoda notes that there are other possible uses. For example, if smartphones or laptops are equipped with Bayesian machine learning, they won't have to share personal data with large companies to determine the interests of users; calculations can be carried out efficiently inside the device. Autonomous cars can also learn to adapt to their environment much faster using this learning method.
If you teach artificial intelligence to learn on its own, it doesn't have to be on a leash.
ILYA KHEL