Organizations are increasingly looking for ways to capitalize on all of the unstructured data they’ve been collecting and stashing. Deep learning applications could offer an answer
At Twitter Inc., Hugo Larochelle’s job is to develop an understanding of how users of the social network are connected to each other and what interests them in order to categorize and promote content that includes tweets, images and videos. To help accomplish that, he and his fellow data analysts use an emerging technology: deep learning tools.
As Larochelle, a research scientist at Twitter, explained during a presentation at the Deep Learning Summit in Boston this month, deep learning is a category of machine learning that seeks to understand complex problems, such as interpreting images or text-based natural language. He and other proponents say deep learning techniques — which lean heavily on the use of neural networks — are more useful than traditional machine learning when data analytics applications involve unstructured data or require subjective interpretations.
And deep learning is quickly becoming a hot field in the realm of advanced data analytics. Larochelle said significant advances have been made in deep learning technology and processes over the past five years as researchers and enterprise analytics teams worked to find a good use for all the unstructured text, image and video data they’re now compiling.
The availability of robust open source tools has been another key catalyst. Larochelle said his team at Twitter uses Google’s TensorFlow and Torch, an open source machine learning platform developed by researchers at Facebook, Google and Twitter. Such tools have made it easier, he added, fordata scientists and other analysts to build deep learning applications.
“All of these factors combined together have created a perfect storm where deep learning has become very successful in the industry,” Larochelle said.
Big data plays big role in deep learning
Aside from the variety of data types that organizations are storing, the sheer volume of data has been a key factor spurring development of deep learning tools and techniques. Daniel McDuff, director of research and principal scientist at Affectiva, a company based in Waltham, Mass., that’s developing deep learning software for recognizing emotions in facial expressions, said the technology has become useful only recently after the startup accumulated sufficient data to get it to work properly.
Affectiva grew out of a research project at the MIT Media Lab, officially launched in 2009. Initially, the company’s researchers, using videos of a couple hundred volunteers, tried to “train” facial recognition software to interpret the emotional states of people. But at the conference, McDuff said there was too much variation in such a small data set to be reliable. Over the years, Affectiva has built a video library with footage of millions of people. Now, according to McDuff, its machine learning algorithms have become proficient at making more generalized assessments of the emotions people are feeling.
“It’s really not been possible to train these models with the volume of data that was available in the past,” McDuff said. But now, he added, the company’s research team has a “rich source of information,” which has improved the accuracy of its analytical models.
Business case for deep learning grows
While a lot of deep learning projects are still in the research phase, the case for applying the technology to more traditional business problems is growing. Yoshua Bengio, a professor of machine learning at the Université de Montréal, said software based on deep learning is permeating everyday life.