Phil Goldstein is a former web editor of the CDW family of tech magazines and a veteran technology journalist. He lives in Washington, D.C., with his wife and their animals: a dog named Brenna, and two cats, Grady and Princess.
In recent months, the Trump administration and Congress have also indicated their growing support for direct federal funding and R&D. In March, the administration released an Overview of Administration Activities in Artificial Intelligence, and on May 8, the AI in Government Act was reintroduced by lawmakers aiming to increase federal AI talent and establish AI governance plans for agencies.
Deep learning is one of the facets of machine learning and AI that agencies are actively deploying, and it holds wide-ranging potential for the government. Deep learning technology can be used for an array of applications that have benefits for both civilian and defense or intelligence agencies.
What Is Deep Learning?
To understand what deep learning is, it’s necessary to get a quick primer on AI and machine learning.
“In a nutshell, AI is software that can do something it wasn’t explicitly written to do,” Tom Leinberger, business development manager for high performance computing, financial service at CDW, writes in a blog post comparing machine learning vs. artificial intelligence. “That capability is enabled by machine learning (ML), which uses linear regression, decision trees and other algorithms to identify patterns. ML teaches the AI which patterns to look for so it can respond when it sees them.”
Within machine learning is the smaller subcategory of deep learning, “which has been used recently to make advances in a variety of domains which previously were too challenging to address,” says Susie Adams, CTO of Microsoft Federal.
“Deep Learning, while very advanced, is also the most resource-intensive, needing thousands of times the amount of data and computation required for simpler techniques, but also providing human-level performance on an entirely new class of problems such as object detection, handwriting decoding and so on,” Adams says.
Deep learning takes advantage of “algorithms that learn by using a large, many-layered collection of connected processes and exposing these processors to a vast set of examples,” Adams adds.
These layered algorithms, known as called artificial neural networks, are inspired by the biological neural network that the human brain uses, Adams says. “While a standard machine learning model would need to be told how it should make an accurate prediction (by feeding it more data), a deep learning model is able to learn on its own,” she says.
Deep learning enables numerous applications, including video analytics, natural language processing and language recognition, computer vision, and image colorization. Adams notes that agencies can potentially use deep learning for everything from fraud detection to handwriting recognition, image search and recognition, natural language translation, and predictive maintenance.
Adams says that the real benefits of deep learning will be realized “when agencies begin implementing solutions that digitally transform how government meets its mission.”
For example, they could potentially use real-time language translation to help citizens with services, incorporate video and image recognition into scenarios that would typically require humans to visually inspect each document, and use deep learning to predict when a piece of military equipment needs to be worked on before it fails.
Here is a breakdown of some of the key deep learning applications being used in government.
As a DHS workshop outline explains, video analytics applications use “information and knowledge from video data content to address a particular applied information processing need.” Video analytics can automate the time-consuming task of monitoring live streams of video and helps speed up the process of searching through video footage.
Video analytics typically helps agencies answer what are commonly referred to as the “w” questions, the outline notes:
Who (people detection and identification)
What (object, activity, event, behavior and relationship analysis)
Where (frame space, 3D space and world map space)
When (date, time of day, time of year)
The National Institute of Standards and Technology said earlier this year that it wanted video of a simulated robbery to determine if video analytics technology could identify a weapon in the footage.
“The collected video data shall allow measurement of the effects of variances in video quality on the performance of evolving video analytics in a simulated public safety setting,” NIST says in the solicitation. “The intent is to drive advances in computer vision that ultimately will directly benefit the public safety user community.”
Similarly, as Nextgov reports, IARPA is now “recruiting teams to build bigger, better datasets to train computer vision algorithms that would monitor people as they move through urban environments. The training data would improve the tech’s ability to link together footage from a large network of security cameras, allowing it to better track and identify potential targets.”
As Deloitte’s William Eggers, Matt Gracie and Neha Malik write in an article for Deloitte Insights, natural language processing tools “encompass the entire cycle of recognizing human speech, understanding and processing natural language, and generating text that can be read and interpreted by humans.”
“With recent technological advances, computers now can read, understand, and use human language,” the Deloitte experts say. “They can even measure the sentiment behind certain text and speech. These capabilities allow government agencies to recognize patterns, categorize topics, and analyze public opinion.”
Earlier this year, DARPA unveiled a new program known as Grounded Artificial Intelligence Language Acquisition, to “build an automated language acquisition system that learns language the way children do — extracting meaning from hearing sounds while observing the environment,” as GCN reports.
Computer vision is made up of several technologies that are combined to form a new kind of AI tool. Ultimately, it is a method for acquiring, processing and analyzing images, and can automate, through machine learning techniques, what human visual analysis can perform.
One way to imagine computer vision technology is as a stool with three legs: sensing hardware, software (algorithms, specifically) and the data sets they produce when combined.
While a standard machine learning model would need to be told how it should make an accurate prediction (by feeding it more data), a deep learning model is able to learn on its own.”
Susie Adams
CTO, Microsoft Federal
Since machine learning governs computers rather than rules-based engines, with each processing of an image by the algorithms that underpin computer vision platforms, the computer refines its techniques and improves its ability to recognize an object.
The National Geospatial-Intelligence Agency has been an enthusiastic proponent of computer vision. NGA says in a technology focus areas document that its analysts need computer vision tools to “preprocess imagery, automatically identify features and objects, and assess change in remotely sensed overhead imagery and full-motion video. The tools should be capable of running multiple algorithms and analytics that are trainable and tunable by users with optional support from engineers and operate within GEOINT exploitation tools, such as electronic light tables and other imagery exploitation programs.”
Last year, the Transportation Security Administration and DHS’s Science and Technology Directorate released a solicitation for new and innovative technologies to enhance security screening at airports. The solicitation, under S&T’s Silicon Valley Innovation Program, is called Object Recognition and Adaptive Algorithms in Passenger Property Screening. In a procurement document, TSA noted that its goal is to “automate the detection decision for all threat items to the greatest extent possible through the application of artificial intelligence techniques.”
Image Colorization in Deep Learning
NGA is also using another form of deep learning known as image colorization. Last year, the agency announceda $15,000 prize for an algorithm that would be able to “rapidly, accurately and automatically colorize large panchromatic images of up to 10 gigapixels in under 10 hours with no human intervention.”
“This goes well beyond simple colorization of an old black and white image in terms of the size of the image and the speed at which the colorization must be accomplished,” the challenge announcement said.
So-called “Colored Earth” images, such as those captured by satellites equipped with multispectral sensors, are generally grainy, less detailed and of lower quality than images collected by satellites with panchromatic grayscale sensors, Linn Wisner, NGA’s Rapid Image Colorization challenge owner, said in a press release. However, the lack of color in panchromatic grayscale images can sometimes be a problem.
“The addition of color makes the images more useful and can assist in assessments of the information within the image,” Wisner said. “Ideally, the best images would combine high resolution with color, and that is the idea behind the challenge.”
