AnalyticsCube

DATA NOW STREAM from daily life: from phones and credit cards and televisions and computers; from the infrastructure of cities; from sensor-equipped buildings, trains, buses, planes, bridges, and factories. The data flow so fast that the total accumulation of the past two years—a zettabyte—dwarfs the prior record of human civilization. “There is a big data revolution,” saysWeatherhead University Professor Gary King. But it is not the quantity of data that is revolutionary. “The big data revolution is that now we can do something with the data.”

The revolution lies in improved statistical and computational methods, not in the exponential growth of storage or even computational capacity, King explains. The doubling of computing power every 18 months (Moore’s Law) “is nothing compared to a big algorithm”—a set of rules that can be used to solve a problem a thousand times faster than conventional computational methods could. One colleague, faced with a mountain of data, figured out that he would need a $2-million computer to analyze it. Instead, King and his graduate students came up with an algorithm within two hours that would do the same thing in 20 minutes—on a laptop: a simple example, but illustrative.

New ways of linking datasets have played a large role in generating new insights. And creative approaches to visualizing data—humans are far better than computers at seeing patterns—frequently prove integral to the process of creating knowledge. Many of the tools now being developed can be used across disciplines as seemingly disparate as astronomy and medicine. Among students, there is a huge appetite for the new field. A Harvard course in data science last fall attracted 400 students, from the schools of law, business, government, design, and medicine, as well from the College, the School of Engineering and Applied Sciences (SEAS), and even MIT. Faculty members have taken note: the Harvard School of Public Health (HSPH) will introduce a new master’s program in computational biology and quantitative genetics next year, likely a precursor to a Ph.D. program. In SEAS, there is talk of organizing a master’s in data science.

“There is a movement of quantification rumbling across fields in academia and science, industry and government and nonprofits,” says King, who directs Harvard’sInstitute for Quantitative Social Science (IQSS), a hub of expertise for interdisciplinary projects aimed at solving problems in human society. Among faculty colleagues, he reports, “Half the members of the government department are doing some type of data analysis, along with much of the sociology department and a good fraction of economics, more than half of the School of Public Health, and a lot in the Medical School.” Even law has been seized by the movement to empirical research—“which is social science,” he says. “It is hard to find an area that hasn’t been affected.”

The story follows a similar pattern in every field, King asserts. The leaders are qualitative experts in their field. Then a statistical researcher who doesn’t know the details of the field comes in and, using modern data analysis, adds tremendous insight and value. As an example, he describes how Kevin Quinn, formerly an assistant professor of government at Harvard, ran a contest comparing his statistical model to the qualitative judgments of 87 law professors to see which could best predict the outcome of all the Supreme Court cases in a year. “The law professors knew the jurisprudence and what each of the justices had decided in previous cases, they knew the case law and all the arguments,” King recalls. “Quinn and his collaborator, Andrew Martin [then an associate professor of political science at Washington University], collected six crude variables on a whole lot of previous cases and did an analysis.” King pauses a moment. “I think you know how this is going to end. It was no contest.” Whenever sufficient information can be quantified, modern statistical methods will outperform an individual or small group of people every time.

In marketing, familiar uses of big data include “recommendation engines” like those used by companies such as Netflix and Amazon to make purchase suggestions based on the prior interests of one customer as compared to millions of others. Target famously (or infamously) used an algorithm to detect when women were pregnant by tracking purchases of items such as unscented lotions—and offered special discounts and coupons to those valuable patrons. Credit-card companies have found unusual associations in the course of mining data to evaluate the risk of default: people who buy anti-scuff pads for their furniture, for example, are highly likely to make their payments.

In the public realm, there are all kinds of applications: allocating police resources by predicting where and when crimes are most likely to occur; finding associations between air quality and health; or using genomic analysis to speed the breeding of crops like rice for drought resistance. In more specialized research, to take one example, creating tools to analyze huge datasets in the biological sciences enabled associate professor of organismic and evolutionary biology Pardis Sabeti, studying the human genome’s billions of base pairs, to identify genes that rose to prominence quickly in the course of human evolution, determining traits such as the ability to digest cow’s milk, or resistance to diseases like malaria.

King himself recently developed a tool for analyzing social media texts. “There are now a billion social-media posts every two days…which represent the largest increase in the capacity of the human race to express itself at any time in the history of the world,” he says. No single person can make sense of what a billion other people are saying. But statistical methods developed by King and his students, who tested his tool on Chinese-language posts, now make that possible. (To learn what he accidentally uncovered about Chinese government censorship practices, see“Reverse-engineering Chinese Censorship.”)

King also designed and implemented “what has been called the largest single experimental design to evaluate a social program in the world, ever,” reports Julio Frenk, dean of HSPH. “My entire career has been guided by the fundamental belief that scientifically derived evidence is the most powerful instrument we have to design enlightened policy and produce a positive social transformation,” says Frenk, who was at the time minister of health for Mexico. When he took office in 2000, more than half that nation’s health expenditures were being paid out of pocket—and each year, four million families were being ruined by catastrophic healthcare expenses. Frenk led a healthcare reform that created, implemented, and then evaluated a new public insurance scheme, Seguro Popular. A requirement to evaluate the program (which he says was projected to cost 1 percent of the GDP of the twelfth-largest economy in the world) was built into the law. So Frenk (with no inkling he would ever come to Harvard), hired “the top person in the world” to conduct the evaluation, Gary King.

Given the complications of running an experiment while the program was in progress, King had to invent new methods for analyzing it. Frenk calls it “great academic work. Seguro Popular has been studied and emulated in dozens of countries around the world thanks to a large extent to the fact that it had this very rigorous research with big data behind it.” King crafted “an incredibly original design,” Frenk explains. Because King compared communities that received public insurance in the first stage (the rollout lasted seven years) to demographically similar communities that hadn’t, the results were “very strong,” Frenk says: any observed effect would be attributable to the program. After just 10 months, King’s study showed that Seguro Popular successfully protected families from catastrophic expenditures due to serious illness, and his work provided guidance for needed improvements, such as public outreach to promote the use of preventive care.

King himself says big data’s potential benefits to society go far beyond what has been accomplished so far. Google has analyzed clusters of search terms by region in the United States to predict flu outbreaks faster than was possible using hospital admission records. “That was a nice demonstration project,” says King, “but it is a tiny fraction of what could be done” if it were possible for academic researchers to access the information held by companies. (Businesses now possess more social-science data than academics do, he notes—a shift from the recent past, when just the opposite was true.) If social scientists could use that material, he says, “We could solve all kinds of problems.” But even in academia, King reports, data are not being shared in many fields. “There are even studies at this university in which you can’t analyze the data unless you make the original collectors of the data co-authors.”

The potential for doing good is perhaps nowhere greater than in public health and medicine, fields in which, King says, “People are literally dying every day” simply because data are not being shared.

Bridges to Business

NATHAN EAGLE, an adjunct assistant professor at HSPH, was one of the first people to mine unstructured data from businesses with an eye to improving public health in the world’s poorest nations. A self-described engineer and “not much of an academic” (despite having held professorships at numerous institutions including MIT), much of his work has focused on innovative uses of cell-phone data. Drawn by the explosive growth of the mobile market in Africa, he moved in 2007 to a rural village on the Kenyan coast and began searching for ways to improve the lives of the people there. Within months, realizing that he would be more effective sharing his skills with others, he began teaching mobile-application development to students in the University of Nairobi’s computer-science department.

While there, he began working with the Kenyan ministry of health on a blood-bank monitoring system. The plan was to recruit nurses across the country to text the current blood-supply levels in their local hospitals to a central database. “We built this beautiful visualization to let the guys at the centralized blood banks in Kenya see in real time what the blood levels were in these rural hospitals,” he explains, “and more importantly, where the blood was needed.” In the first week, it was a giant success, as the nurses texted in the data and central monitors logged in every hour to see where they should replenish the blood supply. “But in the second week, half the nurses stopped texting in the data, and within about a month virtually no nurses were participating anymore.”

Eagle shares this tale of failure because the episode was a valuable learning experience. “The technical implementation was bulletproof,” he says. “It failed because of a fundamental lack of insight on my part…that had to do with the price of a text message. What I failed to appreciate was that an SMS represents a fairly substantial fraction of a rural nurse’s day wage. By asking them to send that text message we were asking them to essentially take a pay cut.”

Fortunately, Eagle was in a position to save the program. Because he was already working with most of the mobile operators in East Africa, he had access to their billing systems. The addition of a simple script let him credit the rural nurses with a small denomination of prepaid air time, about 10 cents’ worth—enough to cover the cost of the SMS “plus about a penny to say thank you in exchange for a properly formatted text message. Virtually every rural nurse reengaged,” he reports, and the program became a “relatively successful endeavor”—leading him to believe that cell phones could “really make an impact” on public health in developing nations, where there is a dearth of data and almost no capacity for disease surveillance.

Eagle’s next project, based in Rwanda, was more ambitious, and it also provided a lesson in one of the pitfalls of working with big data: that it is possible to findcorrelations in very large linked datasets without understanding causation. Working with mobile-phone records (which include the time and location of every call), he began creating models of people’s daily and weekly commuting patterns, termed their “radius of generation.” He began to notice patterns. Abruptly, people in a particular village would stop moving as much; he hypothesized that these patterns might indicate the onset of a communicable disease like the flu. Working with the Rwandan ministry of health, he compared the data on cholera outbreaks to his radius of generation data. Once linked, the two datasets proved startlingly powerful; the radius of generation in a village dropped two full weeks before a cholera outbreak. “We could even predict the magnitude of the outbreak based on the amount of decrease in the radius of generation,” he recalls. “I had built something that was performing in this unbelievable way.”

And in fact it was unbelievable. He tells this story as a “good example of why engineers like myself shouldn’t be doing epidemiology in isolation—and why I ended up joining the School of Public Health rather than staying within a physical-science department.” The model was not predicting cholera outbreaks, but pinpointing floods. “When a village floods and roads wash away, suddenly the radius of generation decreases,” he explains. “And it also makes the village more susceptible in the short term to a cholera outbreak. Ultimately, all this analysis with supercomputers was identifying where there was flooding—data that, frankly, you can get in a lot of other ways.”

Despite this setback, Eagle saw what was missing. If he could couple the data he had from the ministry of health and the mobile operators with on-the-ground reports of what was happening, then he would have a powerful tool for remote disease surveillance. “It opened my eyes to the fact that big data alone can’t solve this type of problem. We had petabytes* of data and yet we were building models that were fundamentally flawed because we didn’t have real insight about what was happening” in remote villages. Eagle has now built a platform that enables him to survey individuals in such countries by paying them small denominations of airtime (as with the Kenyan nurses) in exchange for answering questions: are they experiencing flu-like symptoms, sleeping under bednets, or taking anti-malarials? This ability to gather and link self-reported information to larger datasets has proven a powerful tool—and the survey technology has become a successful commercial entity named Jana, of which Eagle is co-founder and CEO.

New Paradigms—and Pitfalls

WILLY SHIH, Cizik professor of management practice at Harvard Business School, says that one of the most important changes wrought by big data is that their use involves a “fundamentally different way of doing experimental design.” Historically, social scientists would plan an experiment, decide what data to collect, and analyze the data. Now the low cost of storage (“The price of storing a bit of information has dropped 60 percent a year for six decades,” says Shih) has caused a rethinking, as people “collect everything and then search for significant patterns in the data.”

“This approach has risks,” Shih points out. One of the most prominent is data dredging, which involves searching for patterns in huge datasets. A traditional social-science study might assert that the results are significant with 95 percent confidence. That means, Shih points out, “that in one out of 20 instances” when dredging for results, “you will get results that are statistically significant purely by chance. So you have to remember that.” Although this is true for any statistical finding, the enormous number of potential correlations in very large datasets substantially magnifies the risk of finding spurious correlations.

Eagle agrees that “you don’t get good scientific output from throwing everything against the wall and seeing what sticks.” No matter how much data exists, researchers still need to ask the right questions to create a hypothesis, design a test, and use the data to determine whether that hypothesis is true. He sees two looming challenges in data science. First, there aren’t enough people comfortable dealing with petabytes of data. “These skill sets need to get out of the computer-science departments and into public health, social science, and public policy,” he says. “Big data is having a transformative impact across virtually all academic disciplines—it is time for data science to be integrated into the foundational courses for all undergraduates.”

Safeguarding data is his other major concern, because “the privacy implications are profound.” Typically, the owners of huge datasets are very nervous about sharing even anonymized, population-level information like the call records Eagle uses. For the companies that hold it, he says, “There is a lot of downside to making this data open to researchers. We need to figure out ways to mitigate that concern and craft data-usage policies in ways that make these large organizations more comfortable with sharing these data, which ultimately could improve the lives of the millions of people who are generating it—and the societies in which they are living.”

John Quackenbush, an HSPH professor of computational biology and bioinformatics, shares Eagle’s twin concerns. But in some realms of biomedical big data, he says, the privacy problem is not easily addressed. “As soon as you touch genomic data, that information is fundamentally identifiable,” he explains. “I can erase your address and Social Security number and every other identifier, but I can’t anonymize your genome without wiping out the information that I need to analyze.” Privacy in such cases is achieved not through anonymity but by consent paired with data security: granting access only to authorized researchers. Quackenbush is currently collaborating with a dozen investigators—from HSPH, the Dana-Farber Cancer Institute, and a group from MIT’s Lincoln Labs expert in security—to develop methods to address a wide range of biomedical research problems using big data, including privacy.

He is also leading the development of HSPH’s new master’s program in computational biology and quantitative genetics, which is designed to address the extraordinary complexity of biomedical data. As Quackenbush puts it, “You are not just you. You have all this associated health and exposure information that I need in order to interpret your genomic information.”

A primary goal, therefore, is to give students practical skills in the collection, management, analysis, and interpretation of genomic data in the context of all this other health information: electronic medical records, public-health records, Medicare information, and comprehensive-disease data. The program is a joint venture between biostatistics and the department of epidemiology.

Really Big Data

LIKE EAGLE, Quackenbush came to public health from another discipline—in his case, theoretical and high-energy experimental physics. He first began working outside his doctoral field in 1992, when biologists for the Human Genome Project realized they needed people accustomed to collecting, analyzing, managing, and interpreting huge datasets. Physicists have been good at that for a long time.

The first full human genome sequence took five to 15 years to complete, and cost $1 billion to $3 billion (“Depending on whom you ask,” notes Quackenbush). By 2009, eight years later, the cost had dropped to $100,000 and took a year. At that point, says Quackenbush, “if my wife had a rare, difficult cancer, I would have mortgaged our house to sequence her genome.” Now a genome sequence takes a little more than 24 hours and costs about $1,000—the point at which it can be paid for “on a credit card. That simple statement alone,” he says, “underscores why the biomedical sciences have become so data-driven.

“We each carry two copies of the human genome—one from our mother and one from our father—that together comprise 6 billion base pairs,” Quackenbush continues, “a number equivalent to all the seconds in 190 years.” But knowledge of what all the genes encoded in the genome do and how they interact to influence health and disease remains woefully incomplete. To discover that, scientists will have to take genomic data and “put it in the context of your health. And we’ll have to take you and put you in the context of the population in which you live, the environmental factors you are exposed to, and the people you come in contact with—so as we look at the vast amount of data we can generate on you, the only way we can effectively interpret it is to put it in the context of the vast amount of data we can generate on almost everything related to you, your environment, and your health. We are moving from a big data problem to a really big data problem.”

Curtis Huttenhower, an HSPH associate professor of computational biology and bioinformatics, is one of Quackenbush’s really big data collaborators. He studies the function of the human microbiome, the bacteria that live in and on humans, principally in the gut, helping people extract energy from food and maintaining health. “There are 100 times more genes in the bugs than in a human’s genome,” he reports, and “it’s not unusual for someone to share 50 percent or less of their microbes with other people. Because no one has precisely the same combination of gut bacteria, researchers are still learning how those bacteria distinguish us from each other; meanwhile both human and microbial genetic privacy must be maintained.” Not only do microbiome studies confront 100 times more information per human subject than genome studies, that 100 is different from person to person and changes slowly over time with age—and rapidly, as well, in response to factors like diet or antiobiotics. Deep sequencing of 100 people during the human microbiome project, Huttenhower reports, yielded a thousand human genomes’ worth of sequencing data—“and we could have gotten more. But there is still no comprehensive catalog of what affects the microbiome,” says Huttenhower. “We are still learning.”

Recently, he has been studying microbes in the built environment: from the hangstraps of Boston’s transit system to touchscreen machines and human skin. The Sloan Foundation, which funded the project, wants to know what microbes are there and how they got there. Huttenhower is interested in the dynamics of how entire communities of bugs are transferred from one person to another and at what speed. “Everyone tends to have a slightly different version of Helicobacter pylori, a bacterium that can cause gastric cancer and is transmitted vertically from parents to children,” he says. “But what other portions of the microbiome are mostly inherited, rather than acquired from our surroundings? We don’t know yet.” As researchers learn more about how the human genome and the microbiome interact, it might become possible to administer probiotics or more targeted antibiotics to treat or prevent disease. That would represent a tremendous advance in clinical practice because right now, when someone takes a broad spectrum antibiotic, it is “like setting off a nuke,” say Huttenhower. “They instantly change the shape of the microbiome for a few weeks to months.” Exactly how the microbiome recovers is not known.

A major question in microbiome studies involves the dynamics of coevolution: how the bugs evolved in humans over hundreds of thousands of years, and whether changes in the microbiome might be linked to ailments that have become more prevalent recently, such as irritable bowel disease, allergies, and metabolic syndrome (a precursor to diabetes). Because of the timescale of the change in the patterns of these ailments, the causes can’t be genomic, says Huttenhower. “They could be environmental, but the timescale is also right for the kinds of ecological changes that would be needed in microbial communities,” which can change on scales ranging from days to decades.

“Just think about the number of things that have changed in the past 50 years that affect microbes,” he continues. Commercial antibiotics didn’t exist until about 50 years ago; our locations have changed; and over a longer period, we have gone from 75 percent of the population working in agriculture to 2 percent; our exposure to animals has changed; our exposure to the environment; our use of agricultural antibiotics has changed; what we eat has changed; the availability of drugs has changed. There are so many things that are different over that timescale that would specifically affect microbes. That is why there is some weight given to the microbiome link to the hygiene hypothesis”—the theory that lack of early childhood exposure to a diverse microbiota has led to widespread problems in the establishment of healthy immune systems.

Understanding the links between all these effects will involve data analysis that will dwarf the human genome project and become the work of decades. Like Gary King, Huttenhower favors a good algorithm over a big computer when tackling such problems. “We prefer to build models or methods that are efficient enough to run on a[n entry-level] server. But even when you are efficient, when you scale up to populations of hundreds, thousands, or tens of thousands of people,” massive computational capability is needed.

Recently, having realized that large populations of people will need to be studied to advance microbiome science, Huttenhower has begun exploring how to deploy and run his models to Amazon’s cloud—thousands of linked computers running in parallel. Amazon has teamed with the National Institutes of Health to donate server time for such studies. Says Huttenhower, “It’s an important way for getting manageable big data democratized throughout the research community.”

Discerning Patterns in Complexity

MAKING SENSE of the relationships between distinct kinds of information is another challenge facing researchers. What insights can be gleaned from connecting gene sequences, health records, and environmental influences? And how can humans understand the results?

One of the most powerful tools for facilitating understanding of vast datasets is visualization. Hanspeter Pfister, Wang professor of computer science and director of the Institute for Applied Computational Science, works with scientists in genomics and systems biology to help them visualize what are called high-dimensional data sets (with multiple categories of data being compared). For example, members of his group have created a visualization for use by oncologists that connects gene sequence and activation data with cancer types and stages, treatments, and clinical outcomes. That allows the data to be viewed in a way that shows which particular gene expression pattern is associated with high mortality regardless of cancer type, for example, giving an important, actionable insight for how to devise new treatments.

Pfister teaches students how to turn data into visualizations in Computer Science 109, “Data Science,” which he co-teaches with Joseph K. Blitzstein, professor of the practice in statistics. “It is very important to make sure that what we will be presenting to the user is understandable, which means we cannot show it all,” says Pfister. “Aggregation, filtering, and clustering are hugely important for us to reduce the data so that it makes sense for a person to look at.” This is a different method of scientific inquiry that ultimately aims to create systems that let humans combine what they are good at—asking the right questions and interpreting the results—with what machines are good at: computation, analysis, and statistics using large datasets. Student projects have run the gamut from the evolution of the American presidency and the distribution of tweets for competitive product analysis, to predicting the stock market and analyzing the performance of NHL hockey teams.

Pfister’s advanced students and postdoctoral fellows work with scientists who lack the data science skills they now need to conduct their research. “Every collaboration pushes us into some new, unknown territory in computer science,” he says.

The flip side of Pfister’s work in creating visualizations is the automated analysis of images. For example, he works with Knowles professor of molecular and cellular biology Jeff Lichtman, who is also Ramon y Cajal professor of arts and sciences, to reconstruct and visualize neural connections in the brain. Lichtman and his team slice brain tissue very thinly, providing Pfister’s group with stacks of high-resolution images. “Our system then automatically identifies individual cells and labels them consistently,” such that each neuron can be traced through a three-dimensional stack of images, Pfister reports. Even working with only a few hundred neurons involves tens of thousands of connections. One cubic millimeter of mouse brain represents a thousand terabytes (a petabyte) of image data.

Pfister has also worked with radioastronomers. The head teaching fellow in his data science course, astronomer Chris Beaumont, has developed software (Glue) for linking and visualizing large telescope datasets. Beaumont’s former doctoral adviser (for whom he now works as senior software developer on Glue), professor of astronomy Alyssa Goodman, teaches her own course in visualization (Empirical and Mathematical Reasoning 19, “The Art of Numbers”). Goodman uses visualization as an exploratory technique in her efforts to understand interstellar gas—the stuff of which stars are born. “The data volume is not a concern,” she says; even though a big telescope can capture a petabyte of data in a night, astronomers have a long history of dealing with large quantities of data. The trick, she says, is making sense of it all. Data visualizations can lead to new insights, she says, because “humans are much better at pattern recognition” than computers. In a recent presentation, she showed how a three-dimensional visualization of a cloud of gas in interstellar space had led to the discovery of a previously unknown cloud structure. She will often work by moving from a visualization back to math, and then back to another visualization.

Many of the visualization tools that have been created for medical imaging and analysis can be adapted for use in astronomy, she says. A former undergraduate advisee of Goodman’s, Michelle Borkin ’06, now a doctoral candidate in SEAS (Goodman and Pfister are her co-advisers), has explored cross-disciplinary uses of data-visualization techniques, and conducted usability studies of these visualizations. In a particularly successful example, she showed how different ways of displaying blood-flow could dramatically change a cardiac physician’s ability to diagnose heart disease. Collaborating with doctors and simulators in a project to model blood flow called “Multiscale Hemodynamics,” Borkin first tested a color-coded visual representation of blood flow in branched arteries built from billions of blood cells and millions of fluid points. Physicians were able to locate and successfully diagnose arterial blockages only 39 percent of the time. Using Borkin’s novel visualization—akin to a linear side-view of the patient’s arteries—improved the rate of successful diagnosis to 62 percent. Then, simply by changing the colors based on an understanding of the way the human visual cortex works, Borkin found she could raise the rate of successful diagnosis to 91 percent.

Visualization tools even have application in the study of collections, says Pfister.Professor of romance language and literatures Jeffrey Schnapp, faculty director of Harvard’s metaLAB, is currently at work on a system for translating collections metadata into readily comprehensible, information-rich visualizations. Starting with a dataset of 17,000 photographs—trivial by big data standards—from the missing paintings of the Italian Renaissance collection assembled by Bernard Berenson (works that were photographed but have subsequently disappeared), Schnapp and colleagues have created a way to explore the collection by means of the existing descriptions of objects, classifications, provenance data, media, materials, and subject tags.

The traditional use of such inventory data was to locate and track individual objects, he continues. “We are instead creating a platform that you can use to make arguments, and to study collections as aggregates from multiple angles. I can’t look at everything in the Fogg Museum’s collections even if I am Tom Lentz [Cabot director of the Harvard Art Museums], because there are 250,000 objects. Even if I could assemble them all in a single room,” Schnapp says, “I couldn’t possibly see them all.” But with a well-structured dataset, “We can tell stories: about place, time, distribution of media, shifting themes through history and on and on.” In the case of the Berenson photo collection, one might ask, “What sorts of stories does the collection tell us about the market for Renaissance paintings during Berenson’s lifetime? Where are the originals now? Do they still exist? Who took the photographs and why? How did the photo formats evolve with progress in photographic techniques?”

This type of little “big data” project makes the incomprehensible navigable and potentially understandable. “Finding imaginative, innovative solutions for creatingqualitative experiences of collections is the key to making them count,” Schnapp says. Millions of photographs in the collections of institutions such as the Smithsonian, for example, will probably never be catalogued, even though they represent the richest, most complete record of life in America. It might take an archivist half a day just to research a single one, Schnapp points out. But the photographs are being digitized, and as they come on line, ordinary citizens with local information and experience can contribute to making them intelligible in ways that add value to the collection as an aggregate. The Berenson photographs are mostly of secondary works of art, and therefore not necessarily as interesting individually as they are as a collection. They perhaps tell stories about how works were produced in studios, or how they circulated. Visualizations of the collection grouped by subject are telling, if not surprising. Jesus represents the largest portion, then Mary, and so on down to tiny outliers, such as a portrait of a woman holding a book, that raise rich questions for the humanities, even though a computer scientist might regard them as problems to fix. “We’re on the culture side of the divide,” Schnapp says, “so we sometimes view big data from a slightly different angle, in that we are interested in the ability to zoom between the micro level of analysis (an individual object), the macro level (a collection), and the massively macro (multiple collections) to see what new knowledge allows you to expose, and the stories it lets you tell.”

• • •

DATA, IN THE FINAL ANALYSIS, are evidence. The forward edge of science, whether it drives a business or marketing decision, provides an insight into Renaissance painting, or leads to a medical breakthrough, is increasingly being driven by quantities of information that humans can understand only with the help of math and machines. Those who possess the skills to parse this ever-growing trove of information sense that they are making history in many realms of inquiry. “The data themselves, unless they are actionable, aren’t relevant or interesting,” is Nathan Eagle’s view. “What is interesting,” he says, “is what we can now do with them to make people’s lives better.” John Quackenbush says simply: “From Kepler using Tycho Brahe’s data to build a heliocentric model of the solar system, to the birth of statistical quantum mechanics, to Darwin’s theory of evolution, to the modern theory of the gene, every major scientific revolution has been driven by one thing, and that is data.”