Analogies are the comparison of one thing with another, most commonly with the goal of explaining or clarifying a certain concept. Like a well-chosen metaphor, a good analogy can be a great tool for people such as writers. However, it can also be crucial for problem-solving, since comparing separate problems or methods in this way can be used to highlight underlying — often times useful — similarities. For instance, a few years ago a car mechanic was watching a YouTube video showing how to extract a cork from a wine bottle when he struck upon using the same approximate method for helping babies stuck in the birth canal. Unfortunately, analogies are not the most straightforward idea for a computer to understand. As we turn to artificial intelligence to solve more and more of our problems, the need for software that can understand analogies, therefore, becomes more important. That is where a new deep-learning project from Carnegie Mellon University and the Hebrew University of Jerusalem comes into play. AI researchers there have created a means by which smart agents can analyze databases of patents, inventions, and research papers, and identify ideas which could be useful for solving new problems or creating innovative products. “Finding useful analogies automatically is very hard for computers,” Dafna Shahaf, a CMU alumnus and a computer scientist at Hebrew University, told Digital Trends. “Previous work relies heavily on hand-created databases, taking thousands of person-hours to create. Instead, we decided to try the data-driven approach. There are lots of idea repositories online, with millions of problems and solutions. We took advantage of recent advances in deep learning and AI, and found a lightweight way to learn, given a product description, a representation for what the product does, and how it does it. This allows us to ask questions such as ‘find me another product in the dataset that solves a similar problem in a completely different way’ and ‘find me another use for this product.’” This is not necessarily about handing over yet another sphere of human endeavor, though. In a test of the work, Shahaf said that human participants were tasked with problems in need of solving — such as extending the battery of a cell phone. “[The] people who were exposed to inspirations from our algorithm came up with significantly more creative ideas,” she said. “We could even see in some cases how the algorithm helped people explore more diverse parts of the design space — things they would not have thought of on their own.” The researchers will present their work this week at KDD 2017, the Conference on Knowledge Discovery and Data Mining, in Halifax, Nova Scotia.

Read the source article at