4

First Research Session: Brain and Neuroscience

The first research panel was moderated by Valerie Reyna (Cornell University) and showcased cutting-edge work in the area of neuroscience and studies of the brain. Panelists included Paul Glimcher, professor of neural science, economics, and psychology at New York University; Read Montague, professor of physics at Virginia Polytechnic Institute and State University; and Elizabeth Phelps, professor of psychology and neural science at New York University. The panelists presented overviews of their research programs and highlighted key findings, methodologies, data considerations, and relevance to the work of analysts in the intelligence community (IC).

THE KAVLI HUMAN PROJECT

Paul Glimcher focused on the process of data collection as he considered how social and behavioral scientists try to collect sufficient data on the characteristics of individuals for the purposes of making predictions about behaviors. He also considered research gaps and how they might be addressed.

He first introduced the concept of a phenotype, which he defined “as the set of all observable characteristics of an individual that result from the interaction of its genome and its environment.” A set of data on an individual’s genome, environment, and actions provides as complete a description of an individual actor as possible. Generation of phenotypes across a representative population could help characterize an entire population. He argued that limitations in phenotyping are not a problem of technology and analytic capabilities at this point, but more of a failure to collect appropri-

ate phenotyping data. Scientists have access to large datasets, but they end up with very select groups of people who often are not representative of the populations to be characterized. In addition, data are often gathered on people for short periods of time, and as such, the science community lacks sufficient longitudinal information to draw accurate inferences. Glimcher pointed out that well-studied populations tend not to be representative of the nation. For example, a tremendous amount is known about people with certain kinds of diseases. Errors are made when findings from this population are extrapolated back to the national population.

In order to build a database to characterize a general population, Glimcher said three goals need to be achieved: (1) extract a representative sample of people, a perfect cross-section of the community of interest; (2) monitor them for a long time; and (3) collect a wide range of data on them (depth)—for example, their genomes, spending habits, health care, and social networks. Such a project would be expensive (an estimated $1,500–$2,000/person). To make it more manageable, according to Glimcher, developers need to reconsider capturing all the data on everyone in a giant cohort and instead think about building a representative phenotypic group from select small cohorts, representative of defined populations. He suggested that models for behaviors of interest be used as starting points to determine what types of individual data should be collected.

Glimcher acknowledged that most people think wide, meaningful phenotyping is impossible, but he remains optimistic about the opportunities. A large body of survey work, illustrated by Glimcher, has shown that longitudinal studies are possible and people are willing to participate for the greater good. Unfortunately, the individual data currently collected in longitudinal surveys¹ are quite narrow and cannot be linked to data in other longitudinal surveys because of the differences in survey samples.

Glimcher introduced the Kavli HUMAN project² as an effort to develop a rapid phenotyping tool and stable platform to capture data at scale and at low cost of representative subpopulations in the United States. Data are to be collected from the first cohort in New York City in 2017 at a cost of about $12 million. Glimcher noted that members of the public are willing to participate in the project with assurances that their data are maintained for academic use and not given to commercial entities like Google or Amazon or to the government. (The project works with municipal-level government to improve areas like health and education but is insulated from it, he noted.)

___________________

¹ Examples of continuing longitudinal surveys include the U.S. Health and Retirement Survey, the Fragile Families Survey, and the Longitudinal Dynamics Survey.

² The Kavli HUMAN project has developed over the course of 3.5 years with support from New York University and a large group of national advisers distributed across the U.S. academic and corporate spaces, with the financial support of the Kavli Foundation in Santa Barbara.

Glimcher acknowledged that the future of phenotyping will need to depend on giant databases but not with data from a billion people. He said data can be collected from a much smaller sample to give the insight needed. Accumulating data on 100 million people is only important if the event(s) of interest has a base rate of 1 in a million, he said. Generally, according to Glimcher, that level of precision would not be utilized, and the collection cost would be huge. For most events (rates 1 in 1,000, or 1 in 10,000), samples of 10,000 to 100,000 are more cost-effective and retain the ability to observe events with these base rates.

He emphasized that representativeness is a critical feature of any baseline collection operation. It is often not available, which he said has led to false estimations of base rates on a number of fronts (e.g., health conditions). The problem is that baseline datasets are often gathered based on opportunities. He underscored that datasets need to be designed from the start to be representative. He also pointed out that breadth of data is only powerful if integrated at the individual level with the ability to link assets to characteristics and actions.

In closing, Glimcher noted that the Kavli HUMAN project has developed with the goal of generating new understandings in the civilian and academic world of who people are and how agents operate. The project has created novel recruiting and retention methods to make the study both attractive and affordable to potential subjects. Messages of a public service mission are important in getting people to share their data. Glimcher reported that about 40 percent of New Yorkers expressed willingness to participate in these studies, even though the studies are very invasive and require a lot of work. The project will measure a number of items: traditional medical data, ranging from genome to blood chemistry to real-time electronic medical records; detailed real-time financial data about wealth, labor allocation, and taxes; swipe-level data about consumer purchase behaviors; social network data, such as SMS, MMS, telephone, email, browsing, geotracks, and Bluetooth/MAC addresses of local emitters; education; family interactions; environmental data; and criminal justice experiences. He noted that the ability to track data for adults as well as for children and low-functioning elders has been developed. He viewed the project as “a telescope for humanity” that could serve as a tool to move SBS research forward over the course of the next decade or two.

NEW, NEW NEUROSCIENCE

Read Montague commented on the ambition of the neuroscience community to describe neural function from molecules and cells up through the various parts of cognition that impact the behavior of individuals and of composite groups of people. He acknowledged the ambition is easy to

say but hard to do. There are dynamics at different levels and interactions between these levels.

The field has changed its perspective, according to Montague, from looking outside to looking inside. Previously, the brain was modeled as an engineering system that took streams of information from the outside world and synthesized and reassembled them into what ended up as a perception of behavior. Now, he said, the field is modeling what goes on inside the brain, with the premise that the brain houses its own deep templates of the world. An important shift in scientific thinking, according to Montague, is that the intersection of these templates with the world generates data that the brain responds to. In terms of investigating neural functioning, the old approach relied on modeling starting from outside behaviors. The new approach starts with measurement of brain activity in an agnostic way of stimulating and predicting outward responses.

Montague provided two examples of the new types of scientific investigations. In the first experiment, people (subjects) would be shown a series of evocative visual stimuli from the International Affective Picture Systems database. He illustrated with an image similar to the ones used in the database—a picture of a colleague putting a bunch of worms in his mouth. Subjects are shown pictures for a short period of time and brain activity is recorded using a functional MRI (fMRI). Researchers have found, according to Montague, that an individual’s political ideology (as measured by the Wilson-Patterson survey) can be predicted from a record of his or her brain activity on the visual task.³ The research also discovered that the fMRI responses to a subset of the pictures shown (the disgusting ones) were what predicted political orientation (on a conservative end of scale compared to a liberal end of scale). Furthermore, data from asking the subjects to rate pictures consciously were not predictive at all. Montague noted that this shows that the conscious report (behavior) can deviate from what the brain portrays on the current state of fMRI. Research is looking at a subliminal version: that is, what happens if subjects do not even consciously see the pictures if they are flashed for minimal amounts of time.

In the second experiment, an average of 20 people per test group work online at the same time, performing parallel tasks in a market scenario to buy, sell, or hold stocks. The brain activity of three subjects per test group is scanned (about 16 groups tested so far). Montague illustrated that when multiple people perform the task at the same time, bubbles (inflated prices) emerge on various rounds despite a relatively flat mean price. The research

___________________

³ For more information, see Ahn, W-Y., et al. (2014). Nonpolitical images evoke neural predictors of political ideology. Current Biology, 24(22):2693-2699.

found that brain activity in the nucleus accumbens⁴ tracks the bubble of the market in the simulation. Notably, subjects for whom the brain activity predicts their buying behavior tend to do worse in the market scenario than other subjects. Montague suggested that this experiment illustrates situations where what people value is not just a function of themselves, and not even just a function of other people, but also a function of blind mechanisms in the brain and the rules of the marketplace.

EMOTION AND DECISION MAKING

Elizabeth Phelps started her presentation with the recognition that the idea that emotional choices are generally considered irrational and poor, while cognitive choices are considered rational and goal-directed, is a misconception and does not apply to the scientific characterization of decision-making. This view is overly simplistic, according to Phelps, and should be abandoned. Research has shown that the influence (or effect) of emotion on decisions is modulatory; its influence shifts depending on the type of decisions, and the shift can be good or bad depending on circumstance.

Phelps reported that the influence of emotion acts in two specific ways. One is considered integral to the decision process: the emotional reaction to the choice of outcomes becomes incorporated into the value of that choice. This could be a good thing, according to Phelps, as emotion can signal threats in the environment. The second way is considered incidental to the decision process: the emotional response comes from an affective state, such as exuberance, a bad mood, or stress. In general, noted Phelps, it is desired to avoid allowing incidental things unrelated to the choice affect or bias the decision. However, it happens to everyone, such as in making an impulsive purchase.

Phelps focused on the incidental way and the affective state of stress. She defined stress as the body’s response to real or implied threat induced by novel, unpredictable, or uncontrollable situations. Science has found that stress has different effects on different parts of the brain. Stress can make parts of the brain—for example, the striatum, the nucleus accumbens, and the amygdala—work better at times. However, even very mild stress can lead to subtle impairments of the function of the prefrontal cortex. Phelps emphasized that acute stress (short periods of stress) and its effects differ from chronic stress (stress that lasts throughout the day or longer periods).

Her laboratory induces acute stress to test its effects. Subjects’ hands are placed in freezing water for 3 minutes. This leads to a reliable increase in cortisol in about 20 minutes. This type of stressor has nothing to do with

___________________

⁴ Reyna pointed out that the nucleus accumbens has been recognized as the reward system in the brain.

subsequent decision tests. Phelps pointed out that “stress is stress”: that is, the neural hormonal changes that occur with stress are rather uniform with different types of stressors.

Phelps illustrated laboratory versions of three types of decisions: (1) attribution, (2) model-based compared to model-free choice, and (3) sequential or foraging. An example of the first decision type is being cut off in traffic. One usually considers the other driver as rude, which is called making a disposition attribution. Social psychologists, according to Phelps, have known about fundamental attribution error for a long time. People tend to make attribution decisions about others whom they do not know well and those not in their social group. In the same situation, if one considers the other driver as rushed, this is called making a situational attribution. Social psychology, noted Phelps, has shown that situational attributions are not the first instinct; they take a little more cognitive control. More recent experiments measuring brain activity continue to confirm this finding, said Phelps. A series of studies in her laboratory found more activity in the prefrontal cortex when subjects look for situational information to make a situational attribution. Her studies also looked at the role acute stress played and found that when subjects were stressed, they were more likely overall to blame a person’s personality in tested situations and less likely to consider the situation or context. Phelps posited that incidental stress biases one to further underestimate the role of situation in defining behavior.

To understand the next decision type, model-based (or free) choice, Phelps explained the difference between laboratory paradigms known as model-free learning and model-based learning. Model-free learning promotes habitual actions without much forethought or attention (done automatically) by linking rewards to context. These types of actions can become insensitive to changes in outcome value or contingencies once they become habits over time. Model-based learning, on the other hand, enables prospective choice of actions and supports adaptation to changes in the environment.

Phelps’s laboratory has developed a two-step task of picking between two images to test the effects of stress on these learning paradigms. In her experiment, the probability that subjects receive a reward after their two selections slowly varies over the course of participation in several trials of the two-step task. If subjects act model-free, they are paying attention to the local context in the task. If subjects act model-based, they are switching based on the transition structure. Phelps notes that this research has discovered that most people follow a combination of both paradigms, with variation in how they use these different types of feedback structures. In this experiment, subjects sometimes performed the two-step task under acute stress. In addition, working memory capacity was also measured for subjects, and it seems to explain some of the differences observed. Phelps reported that

the research found that stress was shifting (or biasing) people to make more model-free choices. However, she said, “this is only true for individuals who on average across [the experiment] group have lower baseline working memory capacity. Individuals with high working memory capacity don’t show this effect.” She concluded that working memory capacity may protect against making more automatic, less goal-directed decision in this case of very mild stress in the laboratory.

For the last type of sequential decisions, Phelps explained that many decisions are made not just one time but repeated until one decides to continue engaging in the line of behavior or switch. The idea of foraging comes from behavioral ecology. Phelps illustrated this with an example of bees drawing nectar from the same flower repeatedly before taking time to move to a new flower as the reward from a current flower decreases. Humans have the same type of continuation decisions, such as in web searches, dating decisions, and jobs.

Phelps defined the optimal switching decision as one where the instantaneous reward rate of staying falls below the average reward rate enough that the cost of time to switch becomes manageable. In her laboratory, a foraging task was simulated by asking subjects to collect as many apples as they could within a fixed amount of time. During a trial, subjects had to decide when it was optimal to use up time to move to another patch of apples. Some of the trials were done under acute stress. Additionally, subjects were asked about their perceived level of stress over the past month as part of the experiment. Phelps reported that the research found that both changes in cortisol to acute stress and subjective perceived stress predicted more of the deviation from optimal switching. Phelps pointed out that humans have learned adaptive behaviors in response to stressors in the environment (e.g., threats to resources); however, when a stressor is completely unrelated to the choice, it is maladaptive. In her laboratory experiments, the unintended consequences influenced by stress were generally considered as negative.

Phelps concluded with several still-unanswered questions, noting that “it’s one thing to identify the relationship between emotion and decision making, another thing to suggest how you change it.” Can resilience to stress be managed to reduce the detrimental effects of incidental stress on decisions? How do the decisions about risk differ in relation to the different types of emotions? Under what conditions might the impact of emotions on decisions be valuable as opposed to detrimental?

DISCUSSION

Charles Gaukel (National Intelligence Council) tried to relate considerations of intelligence analysts to the research presented. He noted that

the issue of representativeness, raised by Glimcher, is very important but perhaps underappreciated by the IC. According to Gaukel, the IC grapples with how to get the attention of policy makers, especially when alerting them to potential emerging risks or opportunities. He said research on effective communications, notably with findings relevant to written communications, would be of value to the IC.

Glimcher noted that academics have worked hard to separate the concepts of risk and ambiguity in the scientific literature. Risk is how people interpret known probabilities of events, and ambiguity applies to situations in which probabilities of occurrence are unknown or partially known. Of note, according to Glimcher, individual traits with regard to risk are not terribly well correlated with the traits regarding ambiguity. He recognized the growing evidence of neurobiological separations between the representations of risk and ambiguity. Adolescents were classically viewed as risk tolerant, but now, with separating the concepts, research finds that adolescents are actually quite risk averse but can tolerate ambiguous situations. Glimcher suggested that adolescents convert their uncertainties to knowledge of risks over the course of adolescence. He added that tolerance for ambiguity tends to decline over one’s life span, and negative life events seem to have a significant role in this decline.

Reyna pointed out that the brain continues to develop much later into one’s life than was previously thought. She raised the question about what differences age and brain development have on decision-making and what this means, in the same context, for decision makers of different ages.

Margaret Polski (George Mason University) asked the panelists if they have looked at developments in mathematics and statistics in creating synthetic data. Montague said he had used generative models to create data for simple social exchange paradigms of two or three people to look at disease categories or psychopathologies. Polski also inquired how occurrences in the physical environment and social context are considered in studies investigating brain function and behavior. Montague noted that models try to account for some innate behavioral wisdom (i.e., instinctual knowledge of what it takes to be a human being) related to one’s physical environment and social context.

Steven Rieber (IARPA) said he saw several challenges in the relevance of large-scale collection of phenotype data to the IC. He said it would be difficult to collect data on populations of interest to the IC because (1) they are non-U.S. people and (2) people of interest tend to be unusual, either in terms of the power they hold or the threat that they pose to the United States. Glimcher clarified that the large-scale phenotyping projects will be noninvasive. The academic community will examine U.S. populations, he noted, to inform health care and other areas of citizen interest. Current research has demonstrated that deep phenotyping is possible, which implies

that phenotyping outside the United States is possible, albeit with both similar and different obstacles.

Jytte Klausen (Brandeis University) noted that her work has focused on behavioral indicators of extreme violence and asked Phelps whether the classification of positive emotions as drivers for action has any relevance. She suggested that perhaps positive emotions are driving this extreme behavior as opposed to what is usually thought of as alienation from society and negative emotions. She then asked about the possibility to create a predictive model for violent extremism given the absence of population markers for anticipating who is more likely to be a violent extremist. Phelps agreed that both positive and negative emotions can drive choices; however, to date, only a crude understanding of the influence of emotions is known. It is harder to generate a positive effect than negative effect in the laboratory, according to Phelps, so the former is probably underinvestigated. She noted that research is just starting to break down affect into different components and look at specific decision processes to uncover how different affect factors, like stress, mood, physiological arousal, and subjective states, come together to influence a choice. On the second question, Phelps explained that it would be very hard to identify indicators of violent extremism without a lot of data from larger populations, looking at individual variability across life span and social and emotional environments.

Glimcher pointed out that a tremendous amount of data can be collected on a known bad actor; however, understanding how that bad actor differs from all the other actors is what is important. The challenge is having a good understanding of all the things necessary across a population to determine which variables are independent and which variables are highly correlated with violent behavior.

Jacqueline Wilson (Civic Fusion International) asked Phelps about any differences on decision-making for those with trauma and longer-term stress. She also asked whether perceptions of risk are different if threats are made through social media or hate speech, as opposed to physical threats. Phelps acknowledged a few findings from the large amount of literature on chronic stress. Chronic stress leads to large changes in the brain and problems with memory. More chronic stress in one’s lifetime increases one’s responsivity. Chronic stress is not the same as acute stress. Phelps underscored that mild acute stress is quite common, and even this mild stress can have an effect on decisions.

Bear Braumoeller (The Ohio State University) pointed to a movement in political science in recent years toward techniques that are more robust to the existence of unmodeled confounders and asked about parallel work in other disciplines. Glimcher recognized the importance of solving the problem of stratifying the variables known to be influential. The challenge, he said, is the covariance structure. If all the variables were independent,

very little data would be needed and the dataset could be modeled, but this is not the case. He emphasized that creating large representative datasets will help identify the covariances. An infinitely large dataset can never be achieved, but he reminded the audience that small representative datasets can be useful and need to be assembled. The good news, according to Glimcher, is that the analytic capabilities exist.