By Bryn Stark, AI Scientist
In advance of the D-Day landings in 1944, the Allies needed on-the-ground pictures of the Normandy beaches in France. In addition to using reconnaissance images, military planners enlisted the BBC and held a vacation photo contest, which netted 10 million photos! Those unsuspecting civilians provided images that helped create a detailed look at the terrain the Allies would encounter.
Imagine the challenges those planners faced in sifting through those millions of photos. How did they accurately sort the content from so many people that arrived in so many formats? Could they verify that a photo labeled Calais wasn’t really Cherbourg?
The planners threw lots of very dedicated people at the mission, using sheer numbers to deal with those challenges. Today, there aren’t enough people in the world to manually tackle the data that’s available.
Thankfully, we now have machine learning, which the government is applying to accelerate mission tasks and do what was previously unachievable. For example, a computer model can be trained to identify people who appear on security cameras and alert personnel to flagged individuals. This dramatically improves security at relatively little cost.
Or, an algorithm can automatically summarize sections of a document, decreasing the time burden of reading through pages of briefs in search of minute details. In each example, machine learning accelerates and augments the work of existing personnel and alleviates the need to hire additional people to achieve time-constrained objectives.
While machine learning provides many benefits, we need to be aware of the limitations of artificial intelligence and other predictive solutions. Machine-learning solutions are largely dependent on high-quality data. Steps must be taken to structure the data as it is created and stored, so that it is readily usable for analysis.
It greatly reduces development costs in analysis. Frequently the most time spent on data science projects involves organizing messy data. Imagine sorting those Normandy photos if none were labeled!
Also pertinent is the common tradeoff between the accuracy of a machine learning model and the interpretability of the process it uses to produce results. Many models that provide the “best” results offer no practical interpretability of their conclusions.
When weighing actions with serious consequences regarding national security, decision-makers select solutions with slight lower accuracy but which maintain interpretability of results to allow human experts to analyze and validate them. We can’t use “the computer said so” as a justification to storm the beaches; we want to know the data sets used, their trustworthiness, and assumptions that were made.
SAIC approaches data science solutions with the mission as the central focus. Keeping in mind the requirements of each customer’s unique problem set, we build and tailor machine-learning solutions. We iterate on possible approaches with a “fail fast” mentality to quickly come to the most suitable solution.
It’s necessary for federal organizations to keep pace technologically with the commercial sector. For this reason, we incorporate and innovate cutting-edge techniques in data science, focusing on tailoring best-in-class approaches and technologies to suit an agency’s particular mission. Continually developing new AI and machine learning methodologies to address national security issues will help the U.S. maintain its competitive edge globally.
About the author: Bryn Stark is an applied mathematician and lead data scientist in the AI lab in SAIC's Solutions and Technology Group. She designs and develops analytics, machine learning, and artificial intelligence solutions to address problem sets across a variety of defense and intelligence missions. Bryn earned a BA in economics from the University of San Francisco and a MS in data science from GalvanizeU, a tech incubator in Silicon Valley.