Data Science newsletter – July 17, 2017

GROUP CURATION: N/A

The charts tell us that San Francisco’s dominance of VC investment over the past five years was unprecedented (or at least, not seen since 1995) — and also that it is fading, fast.

Last year I wrote of San Francisco’s VC investment flood and New York’s smaller but also significant gains in VC market share as a bet on densely populated cities over Silicon Valley-style sprawl. 2 That probably wasn’t all wrong — I’m guessing that at least some of the south-to-north shift in the Bay Area’s center of startup gravity will endure. But the San Francisco VC boom is also increasingly looking like it might be something else: a bubble that has begun to deflate.

This near-death experience has not only changed my life and priorities, but also altered my view of artificial intelligence—the field that captured my selfish attention for all those years. This personal reformation gave me an enlightened view of what AI should mean for humanity. Many of the recent discussions about AI have concluded that this scientific advance will likely take over the world, dominate humans, and end poorly for mankind. But my near-death experience has enabled me to envision an alternate ending to the AI story—one that makes the most of this amazing technology while empowering humans not just to survive, but to thrive.

Daniela Rus’s morning is packed. My arrival appears to come as a bit of a surprise, as she readies herself to enter the gauntlet of wall-to-wall meetings. She considers the situation for a moment before inviting me into her office, where a group of students are already patiently waiting to talk self-driving cars. “You can’t report about any of the findings,” Rus says with a smile. “But you can come in.”

Rus has allowed me to sit in for a packed morning of team meetings. It’s a generous gesture, but more to the point, it’s the only way to manage some face-to-face time with the head of MIT’s groundbreaking Computer Science and Artificial Intelligence Library (otherwise known as CSAIL). It’s a non-stop job, heading up the largest lab on MIT’s Cambridge, Massachusetts campus and, from the looks of it, Rus never rests. “There’s no time for an interview,” she explains, as we settle into the meeting. “Maybe during lunch.”

Patient outcomes are taking over from products and services as the focus of healthcare. But reorienting away from product development toward a holistic approach to patients demands the convergence of data from every part of the healthcare system. In this interview, part of our Biopharma Frontiers series on how the pharmaceutical industry is evolving, Jared Josleyn, global head of corporate development at Alphabet-owned Verily Life Sciences, talks with McKinsey’s Michele Raviscioni about the need to integrate health data and apply it to patients’ lives in ways that achieve enduring impact.

Holdren is now back at Harvard University, where he is a professor of environmental policy in both the John F. Kennedy School of Government and the Department of Earth and Planetary Sciences. He says he is troubled by what has happened to his office, and to science policy, under Trump. Holdren spoke with ScienceInsider about those concerns and about the role OSTP plays in supporting the president’s agenda.

Can you imagine watching 20,000 videos, 16 minutes apiece, of fruit flies walking, grooming, and chasing mates? Fortunately, you don’t have to, because scientists have designed a computer program that can do it faster. Aided by artificial intelligence, researchers have made 100 billion annotations of behavior from 400,000 flies to create a collection of maps linking fly mannerisms to their corresponding brain regions.

Experts say the work is a significant step toward understanding how both simple and complex behaviors can be tied to specific circuits in the brain.

Today, you can watch a 30-minute deep learning tutorial online, spin up a 10-node cluster over the weekend to experiment, and shut it down on Monday when you’re done – all for the cost of a few hundred bucks. Betting big on an AI future, cloud providers are investing resources to simplify and promote machine learning to win new cloud customers. This has led to an unprecedented level of accessibility that is breeding grassroots innovation in AI. A comparable technology democratization occurred with the internet in the 1990s and, if AI innovation follows a similar trajectory, the world will be a very interesting place in five years.

First, advances in computing technology (GPU chips and cloud computing, in particular) are enabling engineers to solve problems in ways that weren’t possible before.

IBM scientists reported in the Proceedings of the National Academy of Sciences that they’ve adapted CNNs to run on their TrueNorth chip. Other research groups have also reported progress on the solution. The TrueNorth system matches the accuracy of the best current systems in image and voice recognition, but it uses a small fraction of the energy and operates at many times the speed. Finally, combining convolutional nets with neuromorphic chips could create more than just a jargony mouthful; it could lead to smarter cars and to cellphones that efficiently understand our verbal commands—even when we have our mouths full.

Robots posing as people have become a menace. For popular Broadway shows (need we say “Hamilton”?), it is actually bots, not humans, who do much and maybe most of the ticket buying. Shows sell out immediately, and the middlemen (quite literally, evil robot masters) reap millions in ill-gotten gains.

Philip Howard, who runs the Computational Propaganda Research Project at Oxford, studied the deployment of propaganda bots during voting on Brexit, and the recent American and French presidential elections. Twitter is particularly distorted by its millions of robot accounts; during the French election, it was principally Twitter robots who were trying to make #MacronLeaks into a scandal. Facebook has admitted it was essentially hacked during the American election in November. In Michigan, Mr. Howard notes, “junk news was shared just as widely as professional news in the days leading up to the election.”

A new Rhode Island law is designed to protect draft reports and working papers of researchers at the state’s university and colleges from public disclosure.

The law, enacted late last month, modifies the state’s public records statute that governs public access to information from government entities. Rhode Island may be the first state that has revised its public records law specifically to head off attacks on academics at state institutions of higher education, says Michael Halpern, deputy director of science and democracy at the Union of Concerned Scientists.

“A fast and canonical project setup for TensorFlow models. The most difficult part of getting started with TensorFlow isn’t deep learning, it’s putting together hundreds of API calls into a cohesive model.”

Taking a systems approach to data thinking allows you not only to solve problems more efficiently, but to more deeply understand (and critique) the data machinery that ubiquitously affects our day-to-day lives.

“Sukhoi is built on top of the concept of miners” … “in sukhoi the miners can be placed in structures like lists or dictionaries in order to construct json-like structures for the data thats extracted from the pages.”

Even todays most advanced machine learning models are easily fooled by almost imperceptible perturbations of their inputs. Foolbox is a new Python package to generate such adversarial perturbations and to quantify and compare the robustness of machine learning models. It is build around the idea that the most comparable robustness measure is the minimum perturbation needed to craft an adversarial example. To this end, Foolbox provides reference implementations of most published adversarial attack methods alongside some new ones, all of which perform internal hyperparameter tuning to find the minimum adversarial perturbation. Additionally, Foolbox interfaces with most popular deep learning frameworks such as PyTorch, Keras, TensorFlow, Theano and MXNet, provides a straight forward way to add support for other frameworks and allows different adversarial criteria such as targeted misclassification and top-k misclassification as well as different distance measures.