Deep learning is a type of machine learning that has rapidly emerged to become the cornerstone of many modern artificial intelligence applications. “Deep learning provides a significant improvement in accuracy relative to previous approaches on a wide range of AI tasks, sometimes even exceeding human accuracy,” notes Vivienne Sze, an MIT associate professor and lead instructor of the MIT Professional Education course, Designing Efficient Deep Learning Systems. “This has allowed for the practical use and widespread deployment of AI in a wider variety of applications over the past few years.”

Rapid image recognition and classification, voice recognition, autonomous language translation, and content recommendations, such as those on supplied by search engines and e-commerce sites, are just a few of deep learning's many powerful applications.

Deep learning uses neural networks in successive layers to learn from data in an iterative manner. “Deep learning is especially useful when you’re trying to learn patterns from unstructured data,” says Dan Kirsch, managing director at IT research firm Techstrong Research. “While deep learning is very similar to a traditional neural network, it will have many more hidden layers,” he adds. “The more complex the problem, the more hidden layers there will be in the model.”

Deep learning's power and accuracy stems from its ability to extract high-level features from raw sensory data, using statistical learning on a large amount of data to obtain an effective representation of an input space. “This is different from earlier approaches that used hand-crafted features or rules designed by experts,” Sze says. “In other words, deep learning can learn the relevant features or representations for a given task directly from the data (such as recognizing that wheels belong to a car) without requiring experts to define these features explicitly.”

Data at Work

Deep learning can be applied to a wide range of AI applications, including data analytics to identify patterns, trends, and make predictions, as well as sensing and interfacing with devices, such as smartphones and Internet of Things (IoT) devices. The technology can also be applied to autonomous robots, self-driving cars, and scientific exploration and discovery research, such as protein folding and astronomy, Sze notes.

Vivienne Sze, MIT

Any organization that has access to large datasets and wishes to use that data to help humans or systems make better decisions can, and to a certain extent already are, taking advantage of deep learning. “Early adopters include the financial services sector, as well as media and entertainment and communications companies,” says Michael Scruggs, managing director and applied Intelligence lead at IT consultancy Accenture Federal Services. “Our public sector clients, which include federal agencies, are increasingly leveraging deep learning to improve citizen engagement, decrease fraud, and increase public safety,” he says.

The remarkable thing about deep learning is how prevalent the technology already is in everyday life, observes Christopher Leary, director of insights and data at technology and engineering services firm Sogeti. “Each time we use a virtual assistant, unlock a smartphone using facial recognition, or receive personalized marketing and promotions, we are interacting with a deep learning model.”

Addressing Industry Challenges

Every industry has challenges that can be addressed with deep learning applications, Kirsch says. “It all comes down to the types of challenges and data that you have.”

Wayne Butterfield, director of ISG Automation, a unit of technology research and advisory firm ISG, suggests that enterprises looking to apply deep learning to a specific task should approach the technology with deep pockets, a willingness to experiment, and some common sense. He notes that IT leaders planning a deep learning-based project should always first ask themselves if the final outcome will be “better, faster or cheaper than one from an off-the-shelf product or service.”

Deep learning is already easily accessible to a wide range of organizations. The technology is, in fact, available to virtually any organization that can link an ordinary laptop to a major cloud service provider. “Having access and knowing what problems to solve, or having the skills to solve them, are where the challenges lay,” Butterfield warns. “For most [organizations], taking advantage of solutions that have already utilized deep learning in their product is likely to be the safest, easiest, and cheapest way to take advantage of deep learning, especially at this relatively early stage of its commercial viability.”

Butterfield cautions IT leaders not to try selling deep learning technology itself to business colleagues. “You should sell the solution to a business problem that deep learning is solving,” he says. “It's up to the deep learning experts to ensure they're not selling snake oil, but rather a solution to common business challenges.”

What to Read Next:

Machine Learning Basics Everyone Should Know

Fintech, Cloud, and Bringing Machine Learning to the Edge

AI & Machine Learning: An Enterprise Guide