When an intelligence analyst is presented with photos taken from a satellite or an unmanned aircraft, he is able to look at the images and piece them together with supplemental data to draw conclusions and make decisions.
But along the way, he may form cognitive biases that cloud his judgment and prevent him from considering other explanations for what he is seeing. Being able to understand and predict when and how that bias might interfere with the analytical process — and finding a way to mitigate it — could improve the accuracy and speed of intelligence information that military commanders ultimately rely upon to make critical decisions.
The Intelligence Advanced Research Projects Agency is funding a multi-year program that aims to build a cognitive neuroscience-based model of the sense-making process — the human ability to draw inferences from data that is sparse, noisy and uncertain. Three teams are working to construct a computational model that represents seven major brain systems interacting to produce that human analytical behavior, said project manager Brad Minnery. Though sense-making has been studied by a number of researchers in the psychology, cognitive science, neuroscience and artificial intelligence fields, the current models tend to be descriptive and non-mathematical, he said. “These models are of limited utility to the intelligence community because they cannot make specific, quantitative predictions as to how analysts will perform on complex sense-making tasks where cognitive bias is likely to be a factor,” he explained.
The program, called ICArUS, or integrated cognitive-neural architectures for understanding sense-making, seeks to develop a different kind of model.
“A sense-making model based on cognitive neuroscience will ultimately prove to have more robust predictive ability, especially when it comes to predicting the quirks and idiosyncrasies of human sense-making, including cognitive bias,” Minnery said.
One of the most significant challenges in the program lies in integrating multiple computational models of individual brain systems into a single coherent model, he said. “Historically, most neuroscientists have studied the brain in a piecewise fashion, focusing their efforts on understanding the role that individual brain regions play in individual cognitive functions,” he said. “However, sense-making is a multi-faceted phenomenon that encompasses several cognitive skills and that involves multiple brain areas interacting in concert. Achieving this level of functional integration is essential for program success.”
One of the three teams, led by Lockheed Martin Corp., has assembled a group of researchers from academia, industry and government institutions to help tease out all the pieces to the puzzle.
“Other projects out there are trying to model everything in the brain, down to molecules and individual neurons,” said Jon Darvill, a senior member on the Lockheed Martin Advanced Technology Laboratories engineering staff. But they lack the cognitive side of it — the how and why the brain acts as it does. “We’re bridging the gap between that side and the other,” he said.
The group’s approach is to create a model that uses spiking neurons. Neurons are nerve cells in the brain that process and transmit information electrically and chemically. When they are actively communicating information, they “spike.”
“In order to model biases and perform sense-making the way a human does, you need some model of the dynamics that human brains go through,” said Darvill. Spiking neurons permit a more dynamic model that mimics the humanlike procedure of processing information.
Non-spiking models yield a level firing rate across the board, said Darvill. In the spiking model, scientists can see which clusters of neurons are firing in synchrony, which would indicate concentrated activity that could result in more information.
The teams are halfway through the first of a two-year phase of the program. There are two phases. At the end of each, teams will test their models’ sense-making ability by having them perform a series of tasks involving the analysis of simulated geospatial intelligence data. The models’ performance will be compared to that of human participants performing identical sense-making tasks. Models will be deemed successful if they accurately reproduce human performance patterns, including cognitive biases, said Minnery.
Having predictive models of sense-making will enable the design of new analytic tools and processes that accentuate analysts’ strengths while mitigating weaknesses, he said. For example, an ICArUS-based tool might provide an alert to geospatial intelligence analysts when bias is likely to emerge during an ongoing task.
The models also may lead to the development of automated analysis tools that offload routine sense-making tasks from overburdened human analysts, he added. Analytical software in the future may be able to detect relevant changes in geospatial data — a skill that currently only humans possess.