Data Science newsletter – August 13, 2018

GROUP CURATION: N/A

This Q&A is part of CIFAR’s series on building a research lab.

Graham Taylor is a CIFAR Azrieli Global Scholar in CIFAR’s Learning in Machines & Brains program and an Associate Professor at the School of Engineering, University of Guelph and the Vector Institute. Yoshua Bengio is Co-Director of CIFAR’s Learning in Machines & Brains program and Full Professor in the Department of Computer Science and Operations Research Canada Research Chair in Statistical Learning Algorithms at Université de Montréal.

Facebook would like to have computer science faculty in AI committed to work 80% of their time in industrial jobs and 20% of their time at their university. They call this scheme “co-employment” or “dual affiliation.” This model assumes people can slice their time and attention like a computer, but people can’t do this. Universities and companies are communities, each with their particular missions and values. The values of these communities are often at odds, and researchers must choose where their main commitment lies. By committing researchers to a particular company’s interests, this new model of employment will harm our colleagues, our discipline, and everyone’s future. Like many harms, it comes with benefits for some. But the harm in this proposal outweighs the benefits. If industry wants to support and grow academic computer science, there are much better ways to achieve this.

The proposal will harm our discipline, because it will distract established talent from the special roles of academics: curiosity driven research. Academic scholarship has an excellent record of pursuing ideas into places that are exciting and productive, even if they don’t result in immediate, tangible benefits and especially if they ruffle the feathers of established, powerful institutions. You can’t do that if 80% of your time is spent not annoying a big company. Though big companies belabor promises of complete intellectual freedom to faculty, that can’t and won’t happen because the purpose of companies is to make money for shareholders.

The proposal harms our students directly.

Data is booming. It’s not news to anyone, but the amount of data required by everyone connecting to the Internet is increasing rapidly. This trend is affecting the way data centers are constructed in some unexpected ways, particularly regarding how large data centers will become. Because cloud and hybrid computing are using machine learning and AI and thus require servers to employ tremendous amounts of data, those computing needs are still increasing. Until now, facilities have been getting bigger to meet those needs. But bigger no longer means taller—it can often mean wider.

The generally accepted strategy of stationing monolithic data centers in the largest cities is endangered because despite enormous advances in technology, there’s just too much data. The needs of cities and towns beyond massive population centers are growing, and the businesses and customers located there are moving too much data of their own to be served entirely by large, remote data centers. Instead, the increasing draw of edge data centers is creating a scenario where the number of facilities will increase while the average size will decrease. This is the future of data centers: large-scale computing and small-scale facilities.

Online advertising companies have struggled for several years as Google and Facebook solidified their grip on digital dollars, slowing revenue for the others.

Now, many ad tech companies and their investors are throwing up their hands.

Venture capital money going into ad tech start-ups is falling sharply, helping push a wave of consolidation. Financing reached a high of $2.92 billion in 2015, but this year, it is on pace to be less than half that, according to CB Insights, a financial research firm.

The era of stranger danger is long over: we hop in their cars, we invite them over to assemble our dining room tables, and we even sleep in their beds. You can thank reputation for this new share economy. We know we aren’t willfully entering into a transaction with an axe murderer because when we book Ben’s “Brand New Mini Loft!!” there’s no one-star review that says, “Ben’s freezer was full of dead bodies.”

Though there are obvious benefits, there are drawbacks, too: it’s not hard to imagine (probably because it already exists in China, and in the most recent season of Black Mirror) a slippery slope into a dystopia, where our humanity and character is reduced to a 1-5 rating—an idea we explored in our latest issue.

As part of the resulting piece, I called up Paolo Parigi, a Stanford professor who has worked at Airbnb and Uber studying how the peer-to-peer marketplace is affecting reputation, trust, and our ability to connect with one another. Here, he tells why increased trust in one another may actually lead to increased loneliness.

The question is inaccurate and I will try to explain this through an analogy. Let’s say I go for a walk with a friend and my friend feels a few drops of water that fall on her. She asks me if it’s raining. But as long as there was only few drops of water, it could also be something else.

I tell her that we can get some more relevant information in order to get a more reliable answer. Look at the sky. Are there dark clouds on the sky above? And what does the weather forecast say?

If there are dark clouds above and the weather forecast suggests showers, it’s a safe bet to say it is the start of the rain. The rain always start with a few drops, just the way a climate change starts with a few events.

In the same way as with the observation of the first drops of of water, you could not be sure whether the heatwave is a freak event or the emerging pattern of climate change, if you don’t include other relevant information.

There is a range of different pieces of information which are relevant when it comes to the question about weather events and climate change: (a) statistical evidence, (b) physical processes connecting different aspects, and (c) attribution work.

In a Q&A from Capgemini Consulting on what AI means for large organizations, Michael Schrage, a research fellow at the MIT Sloan School’s Initiative on the Digital Economy, explains how this works in practice. “Companies need to manage data as an asset to get meaningful and measurable economic returns. Start there. Every company is made up of three categories of data managers. Immature managers ask, ‘How do we manage this data?’ More mature managers ask, ‘How do we get value from this data?’ And superlative managers say, ‘What kind of partnerships, governance and technological investments do I need to make to get best-in-class returns on this asset?”

Asking, and certainly answering, the last question requires courage, imagination and creativity. There’s risk involved in that. Get too far out on a limb, and it might break underneath you. People will only take that chance if they feel secure. If their company’s culture not only accepts but encourages ideating on how to get measurable economic returns from data, people will go out on that limb. If they feel there’s no place for big questions and big ideas, people will stay on the ground where they feel safe. But it’s on the ends of limbs that business finds the biggest rewards.

Josh Mitchell’s Defcon presentation analyzes the security of five popular brands of police bodycams (Vievu, Patrol Eyes, Fire Cam, Digital Ally, and CeeSc) and reveals that they are universally terrible, though the Digital Ally models are the least bad of the batch, as Wired’s Lily Hay Newman reports.

All the devices use predictable network addresses that can be used to remotely sense and identify the cameras when they switch on. Attackers could pinpoint intense police activity by watching for groups of cameras that all switch on at the same place and time.

None of the devices use code-signing, making them typical of garbagey Internet of Shit devices. That means that attackers can slip arbitrary code into them. And none of them cryptographically sign the video they take, which would be a relatively strong way of detecting tampering by police officers, their departments, or criminals.

There’s tons more.

“Researchers from the Microsoft Research Montreal Lab, released an open source project called TextWorld, which attempts to train reinforcement learning agents using Text-Based games. The ideas behind TextWorld were captured in a recent research paper published by Microsoft.”

“The World Bank’s new data catalog transforms the way we manage data. It provides access to over 3,000 datasets and 14,000 indicators and includes microdata, time series statistics, and geospatial data.”

WordPress has been around for 15 years. 31.5% of sites use it, and that figure continues to climb. We’re here for the long term, so we need to plan for the long term.

Gutenberg is being built as the base for the next 15 years of WordPress. The first phase, replacing the post editing screen with the new block editor, is getting close to completion. That’s not to say the block editor will stop iterating and improving with WordPress 5.0, rather, this is where we feel confident that we’ve created a foundation that we can build everything else upon.