Data Science newsletter – June 6, 2018

The National Institutes of Health has announced a strategic plan to “modernize the NIH-funded biomedical data ecosystem”. It starts with hiring a Chief Data Strategist who will orchestrate data strategy for the All of Us Program (precision medicine), the BRAIN Initiative (neuroscience), the Cancer Moonshot, and the Opioid Initiative. In essence, this is the role of US biomedical data czar – “By 2025, the total amount of genomics data alone is expected to equal or exceed totals from the three other major producers of large amounts of data: astronomy, YouTube, and Twitter.” (Don’t ask me how Twitter jumped ahead of the rest of Google or Facebook in that quote. They are biomedical scientists, not social media experts.) I’m extremely optimistic that they will strike the appropriate balance between advancing scientific goals as fast as possible without sacrificing patients’ rights. There is a long history of involving biomedical ethicists and developing professional protocols in biomedical research.



NIH released findings from the National Cancer Institute’s Molecular Analysis for Therapy Choice (NCI-MATCH) program. Patient participation is 62.5%, much higher than the predicted 25%. Three drugs used to treat different cancers showed some efficacy, allowing further research to proceed. In all cases, the drugs were matched to patients using recommendations from precision medicine approaches.

Taiwan’s Ministry of Science and Technology forged a new decade-long partnership with GPU-maker NVidia. The extensive partnership includes supercomputing infrastructure, research, education, a start-up accelerator program, and investment into Taiwanese manufacturing, healthcare, urban safety, and transportation. One cool project already underway: Tainan City is using drones to inspect bridges for structural integrity.



Washington State’s Attorney General, Bob Ferguson, is suing Facebook and Google for breaking state law by failing to maintain and make available data about political advertisers on their sites. Neither company has commented. Facebook took in over three times as much as Google, according to a story in TechCrunch.



Canada is fighting hard to prevent brain drain to Silicon Valley and New York with a new article touting all the benefits of staying in Canada. Oddly, they fail to mention the Trudeau–Trump gradient.



US Chief Technology Advisor Michael Kratsios announced that he would do “anything we can do to unlock government data” for artificial intelligence research. He warmly encouraged, “any academic that has any insights” to get in touch regarding which data could be useful. I am hopeful that this will turn into a meaningful commitment, but the fact that he didn’t mention any specific data, timeline, partnership plans, or other clear strategy makes it feel like an empty promise. Let’s hope I am wrong.



Singapore, on the other hand, has released a detailed strategy for weaving ethical considerations into the development of machine learning and other AI technology. All that Singaporean law-making is looking pretty good right now.

Facebook, while under a 2011 consent decree from the US Federal Trade Commission that prohibited the social media behemoth from sharing user data with third parties unless the users had explicitly consented, shared user data with third parties without obtaining their consent. The New York Times investigation found that hardware makers like BlackBerry, Apple, Samsung, Amazon and 56 other device makers were considered “extensions of Facebook” by Facebook executives, thereby exempting them from the FTC consent decree that applied to data sharing partnerships.

If you are feeling confused and angry at Facebook’s disingenuous rule-bending, so is former FTC official Jessica Rich who spoke bluntly in the Times’ article, “Under Facebook’s interpretation, the exception swallows the rule.” At stake is a fundamental, perhaps deliberate, misunderstanding of what consent means. I will clarify how consent works: only first parties can give consent. I can give consent to Samsung to copy and mine my data. I cannot give consent to Samsung to mine data from my Facebook friends.

Still confused? Imagine if we applied Facebook’s version of consent to sexual relationships. It would look a lot like Gayle Rubin’s assessment that when men (usually fathers) were able to give consent to sexual access to women (usually their daughters), women became less than full people. They became valuable physical and symbolic links that protected the economic status of the men in their lives. They were ‘trafficked’ to build kinship ties between men, playing a key role in the patriarchal structure of wealth transfer. A German regulator, who did not cite Gayle Rubin’s discussion of commoditized people, commented that “BlackBerry users seem to have been turned into data dealers, unknowingly and unwillingly.” I won’t claim that using Facebook supports the patriarchy, though at a couple levels of remove, this is true. What I do argue is that corporate powerhouses – device makers and social media behemoths – are enacting an age-old strategy to protect their own dominant positions. Turn people into commodities without their consent. Leave them with no obvious way to refuse, abstain, or see clearly that they should want to do so.



Speaking of data sharing: Two companies – Luna DNA and Nebula Genomics – want to collect DNA and pay people to share it with pharmaceutical companies. Nebula, based in San Francisco, wants pay its customers in cryptocurrency. Neat!



This brings me to a second key requirement when obtaining consent: consent can only be obtained when the person is fully informed of the consequences of their participation. Neither of these companies has explained basic consequences such as: how much users will receive, what would happen if their data were leaked or stolen, and what would happen if the US rescinds its current commitment to the idea that pre-existing conditions cannot be considered when setting health insurance prices. Oh, and I have no idea how to consent someone to being paid in cryptocurrency, given the volatility and short history of that form of payment.



A Canadian company, integrate.ai is going to confront these consent questions immediately as they aim to be an ethical AI company while relying on data sourced from social media sites. They are one to watch – no word yet on how they plan to address the pain point that arises at the overlap between ethical consenting practice, available social media data, and business models relying on the ingestion thereof.



Alphabet shareholders – many of whom are executives at the company – voted against several proposals put forward by employees to link pay to diversity goals and pressure Google to release more information about its content moderation strategies, especially on YouTube. Keep fighting the good fight, Googlers. This isn’t a defeat, it’s just one more step on the path to success.



The day after this shareholder meeting, Alphabet CEO Sundar Pichai announced the underlying principles guiding AI at Google. The fourth principle is “be accountable to people”. I am scratching my head trying to figure out how this is working, given that they literally *just* refused to share information on how content moderation works. [See the government section to see how they are refusing to share political campaign ad data with the state of Washington, who has gone ahead and sued the company.] One of our fellow readers asked me, “will they really not use ML to reinforce stereotypes? It’s literally all their technology currently does! And will they really DESIGN for privacy? It would require discarding their entire codebase and restarting from scratch.” Please, please, let me know what you think. I want to be excited that Google is trying to care – it seems better than Facebook – but capitalism’s demand for growth seems to be the rock against which all good intentions are smashed.

Intel is working on Low Power Display Technology that should improve the battery life in your laptop, possibly giving you up to 28 hours.



Microsoft dropped a data center into the ocean near Scotland. The goal is to use cold ocean water to naturally cool the 864 servers, reducing demand for electricity by up to 20 percent. The data center will be connected to land by a cable that provides both a fiber connection and electrical power. This data center is considered experimental and will be monitored for computational and environmental impact.

Spotify is planning to make a “category defining” hardware product. So far, there’s only speculation on what it will be, but they have registered with the Federal Communication Commission, a key step in the process.

Kipp Johnson (Mount Sinai) Jessica Torres Soto (Stanford), and others who have worked with anchor author Joel Dudley (Mount Sinai School of Medicine) published a paper explaining why and how supervised and unsupervised learning will supercharge doctors treating cardiovascular patients. Because cardiologists often have more data than other types of specialists, they are likely to reap the rewards of data-voracious machine learning methods more quickly. This is a *great* paper to use if you are trying to teach interdisciplinary data science. It walks through various types of machine learning models, compares them to linear regression, and situates them in the context of cardiovascular medicine. Smashingly well done.



University of Manchester researchers have a new behavior-based biometric airport security system. By analyzing the amount of pressure, distance between footsteps, timing of footsteps, and pressure points, the researchers were able to correctly identify individuals with only a 0.7% error rate (I was unable to link to the entire paper to investigate these impressive claims further). The database they used had 20,000 footsteps from 127 people. The researchers are also planning to apply gait analysis to identify the onset of cognitive decline in aging patients. This is the future of security. Expect to see many more behavior based strategies replacing highly exploited, threadbare security protocols revolving around names, social security numbers, dates of birth, and PINs. The future of security is nearly here and it will be so much better.



Ripple, the blockchain company that owns cryptocurrency XRP, has committed $50m to its University Blockchain Research Initiative (URBI) to, you know, study the blockchain. Seventeen universities – including UC-Berkeley, Penn’s Wharton School and Princeton – have already been awarded funding. Because Ripple is a blockchain company, there is some hope it will be able to survive the cryptocurrency hype, but I wouldn’t count on another $50m becoming available annually.



The University of Washington announced a restructuring of the admissions process for Computer Science (CS) majors. In the past, students were admitted first to UW and declared their CS major after a year of coursework. Now, students interested in CS will be admitted directly to the CS major right out of high school. Students who don’t get into the major will have a hard time getting into CS classes because course enrollment policies prioritize students in the major. This is one way to transparently, if bluntly, address the rising demand for CS degrees: let students know they need to commit to the major early and go to a different school if they don’t get in. It would be even better if they could hire more CS faculty, but we all know how hard it is to compete for top talent there.



MIT announced a new joint urban science and computer science major, course 11-6 for those familiar with the Institute’s numbering system. In their words the major will combine, “urban planning and public policy, design and visualization, data analysis, machine learning, and artificial intelligence, pervasive sensor technology, robotics, and other aspects of both computer science and city planning.” My take: It’s a major that focuses on the digitization of the experience of living and working in cities, explicitly considering the human and social impacts of infrastructural change.



The National Academy of Sciences put together a committee to figure out which core competencies undergraduates should develop in data science programs. They came up with a smart, nuanced set of guidelines that emphasize interdiscipinarity: “Data education should expand throughout academic disciplines, but it is necessary that students trained in data science are exposed to continual curricular and co-curricular opportunities to hone vital communication and teamwork skills. Inclusivity refers not only to diversifying student’s exposure to data, but also requires faculty and institutions to think inclusively about collaboration.” Note that they do not think of interdisciplinarity as a set of canonical methods or hard problems, they think of it as a practice, almost as a method in itself. In my experience, this is an area that is severely, cripplingly lacking in data science education. Please, let’s stop thinking of these as (pejorative) soft skills and start thinking of them as professional skills.



Two Swiss scholars, Cristina Kadar and Irena Pletikosa, combined NYC MTA subway data, Foursquare check-in data, American Community Survey (e.g. Census) data, and NYPD open crime data to improve upon current crime forecasting models. They were able to take the current best-in-class R-squared from 65% to 89% by including data about inflows to neighborhoods. This was particularly effective for grand larceny; the Census data was already capable of explaining assault rates fairly well. Of course, what is missing from the underlying data is the comparison between rates of crime committed versus rates of crime reported. Many crimes go unreported, which is likely related to different expectations of apprehension rates. If citizens don’t expect criminals to be prosecuted – say, for instance, women who know their sexual assaults will devolve into “he said it was consensual” or “you shouldn’t have worn/done/said XY or Z” so they stay quiet or communities of color that haven’t been able to rely on police to investigate *homicides* – they are less likely to report crimes in the first place. I’m a bit guilty of blaming the authors for writing the wrong paper, a common academic faux pas, but I am concerned about increasing our ability to ratchet up police pressure on communities that are already over represented in existing datasets. The consequences of overfitting here are serious.