Google’s recent exit from Project Maven put the field of artificial intelligence (AI) in the media limelight once again. Project Maven – a Department of Defense program – aims to use AI to assist in analyzing massive amounts of drone video footage, mostly collected in war zones. While the term AI has roots that go back to the 1950s, the recent focus is really about a specific software technique known as deep learning.
With traditional software development, computer programmers develop algorithms that direct computers to perform highly specific tasks. Deep learning involves training computers to recognize patterns in data via trial and error, using sets of interconnected processors, known as neural networks. These networks work in layers, communicating individual outputs to each other and producing a weighted response that adjusts over time by incorporating feedback about the accuracy or failure of its collective output. As the network sifts through thousands of data points, it is trained to produce ever more accurate results. One of the great advantages brought about by deep learning is that the same neural networks can be used to discover patterns in virtually any type of data.
At CACI, we have applied deep learning to many of our customers’ challenging problems and successfully delivered results that help achieve mission success in areas as diverse as computer vision and natural language processing. There are many difficulties in applying this technique to new problem areas. In order to use a deep learning solution, the first step is usually to acquire labels on tens of thousands of data examples. This labeled data is used to train the deep neural networks to find the important patterns. For many problems, finding adequate data to train these networks is a huge labor challenge, assuming the data exists in the first place. Because of our history in researching and building successful deep learning solutions, CACI teams are experienced at using and creating many different types of deep neural networks to work around data constraints.
One lesson we have learned throughout the years is that deep learning is a journey, not a destination. Advances in this field are very rapid, so committing to a single deep learning model means locking into a capability that is destined to be surpassed by the latest research from academia and the commercial world. Our team develops highly flexible solutions that are designed to incorporate inevitable change. In some cases, we have replaced entire software libraries and changed the deep learning network several times with no impact to the end users.
No technology is useful if it is too complex to be adopted. This is why we bring the science to the user workflow, rather than forcing users to adapt to the science. By making deep learning applications that are as simple to use as a smartphone app, we are helping our customers solve problems more effectively and in far less time than other solutions available today.
By Jasen Halmes, Director of Artificial Intelligence