Data Science newsletter – March 30, 2018

GROUP CURATION: N/A

Objective

Deep learning has become a promising approach for automated support for clinical diagnosis. When medical data samples are limited, collaboration among multiple institutions is necessary to achieve high algorithm performance. However, sharing patient data often has limitations due to technical, legal, or ethical concerns. In this study, we propose methods of distributing deep learning models as an attractive alternative to sharing patient data.
Methods

We simulate the distribution of deep learning models across 4 institutions using various training heuristics and compare the results with a deep learning model trained on centrally hosted patient data. The training heuristics investigated include ensembling single institution models, single weight transfer, and cyclical weight transfer. We evaluated these approaches for image classification in 3 independent image collections (retinal fundus photos, mammography, and ImageNet).
Results

We find that cyclical weight transfer resulted in a performance that was comparable to that of centrally hosted patient data. We also found that there is an improvement in the performance of cyclical weight transfer heuristic with a high frequency of weight transfer.
Conclusions

We show that distributing deep learning models is an effective alternative to sharing patient data. This finding has implications for any collaborative deep learning study.

For almost 20 years, humans have maintained a continuous presence beyond Earth. The International Space Station has provided a habitat where humans can live and work for extended periods of time. Yet, despite having established a permanent base for life in space, terra firma is always in reach — within 254 miles, to be exact. If a crew member were to fall seriously ill, he or she could make the return trip back to Earth in a matter of hours.

“As soon as you venture beyond low Earth orbit, to go to Mars or even further, bailing out no longer is an option,” says Wolfgang Fink, associate professor and Keonjian Endowed Chair in the UA’s College of Engineering. “You’re on your own.”

Fink predicts that in the not-too-far future, humans will work side by side with robotic machines, non-human intelligence and smart devices in ways never seen before. Human logic and thinking will be joined by, and complemented by, artificial brains and reasoning algorithms.

In this post, I’d like to raise awareness about what really worries me when it comes to AI: the highly effective, highly scalable manipulation of human behavior that AI enables, and its malicious use by corporations and governments. Of course, this is not the only tangible risk that arises from the development of cognitive technologies — there are many others, in particular issues related to the harmful biases of machine learning models. Other people are raising awareness of these problems far better than I could. I chose to write about mass population manipulation specifically because I see this risk as pressing and direly under-appreciated.

This risk is already a reality today, and a number of long-term technological trends are going to considerably amplify it over the next few decades. As our lives become increasingly digitized, social media companies get increasing visibility into our lives and minds. At the same time, they gain increasing access to behavioral control vectors — in particular via algorithmic newsfeeds, which control our information consumption. This casts human behavior as an optimization problem, as an AI problem: it becomes possible for social media companies to iteratively tune their control vectors in order to achieve specific behaviors, just like a game AI would iterative refine its play strategy in order to beat a level, driven by score feedback. The only bottleneck to this process is the intelligence of the algorithm in the loop — and as it happens, the largest social network company is currently investing billions in fundamental AI research.

Israel plans to collect data over five years for its digital health project and hopes 100,000 people will volunteer to include their medical records in a database it estimates could bring the country billions of dollars in annual income.

The project will get under way in the fourth quarter and over the next six months the government will work out the mechanisms for collecting the data, Eli Groner, director general of the prime minister’s office (PMO), told reporters on Wednesday.

At our inaugural ACM Future of Computing Academy meeting last June, many of us agreed that the computing research community must do more to address the downsides of our innovations. Indeed, our view is that it is our moral imperative to do so. After several months of discussion, an idea for acting on this imperative began to emerge: we can leverage the gatekeeping functionality of the peer review process.

Below, we describe our specific recommended change to the peer review process. If widely adopted, we believe that this recommendation can make meaningful progress towards a style of computing innovation that is a more unambiguously positive force in the world. We expect that a large proportion of our readers are peer reviewers themselves. If you are a peer reviewer, in many communities, you may be able to try out our recommendation immediately, applying it to the next paper that appears on your review stack and citing this post as justification. In other communities, implementing our recommendation will have to be part of a larger discussion within your community. We hope that you will tell us about these discussions and your experiences implementing our recommendation. This post is part of the FCA Discussions series, which means our recommendations are intended to start a conversation rather than end one!

Two key letters show up throughout Microsoft’s major reorganization announcement this morning: AI.

Artificial intelligence will get even more emphasis as part of the company’s reshuffling of leadership and engineering divisions. In a memo to employees laying out the reorg, Microsoft CEO Satya Nadella called out AI as one of the technologies that will “shape the next phase of innovation.” He continues: “AI capabilities are rapidly advancing across perception and cognition fueled by data and knowledge of the world.”

The company’s AI + Research division, formed two years ago under the leadership of Harry Shum, remains intact, while the revamped cloud division will have a significant AI component as well. AI + Research, or AI + R in Microsoft for short, grew by 60 percent in the first year — from 5,000 people originally to nearly 8,000 people — through hiring and acquisitions, and by bringing aboard additional teams from other parts of the company. Microsoft wouldn’t say how many people work on AI + Research today.

French President Emmanuel Macron will announce his new AI strategy at the Paris-based Collège de France research institute later on Thursday.

Macron has said he does not want France to “miss the AI train” as he introduces measures designed to compete with the United States and China, the current global leaders in AI technology. He has also said he wants to ensure France adopts ethical measures to regulate the industry.

Quietly, stealthily, a new type of neuroscientist is taking shape. From within the myriad ranks of theorists have risen teams of neuroscientists that do science with data on neural activity, on the sparse splutterings of hundreds of neurons. Not the creation of methods for analysing data, though all do that too. Not the gathering of that data, for that requires another, formidable, skill set. But neuroscientists using the full gamut of modern computational techniques on that data to answer scientific questions about the brain. A neural data science has emerged.

Turns out I’m one of them, this clan of neural data scientists. Accidentally. As far as I can tell, that’s how all scientific fields are born: accidentally. Researchers follow their noses, start doing new things, and suddenly find there’s a small crowd of them in the kitchen at parties (because it’s where the drinks are, in the fridge — scientists are smart). So here’s a little manifesto for neural data science: why it’s emerging, and how we might set about doing it.

The why is the same as all areas of science that have spat out a data science: the amount of data is getting out of hand.

Just two months after announcing the beta, Apple is now launching its Apple Health Records feature into the wild, the company announced in a blog post today. The feature will aggregate existing patient-generated data in a user’s Health app with data from their EHR — if the user is a patient at a participating hospital.

In addition to the 12 health systems announced with the beta, 27 more are ready to launch the service, for a total of 39. Anyone with an iPhone and iOS version 11.3 will be able to download the patient-facing side of the feature by updating the Health app in iOS.

Stanford Medicine, Scripps, NYU Langone Medical Center, Partners Health Care, Ochsner Health System in New Orleans, Vanderbilt University Medical Center, and Duke University Medical Center are among the hospitals joining today. Apple previously announced Penn Medicine, Cedars-Sinai in Los Angeles, Johns Hopkins, and Geisinger Health System.

Chemists have a new lab assistant: artificial intelligence. Researchers have developed a ‘deep learning’ computer program that produces blueprints for the sequences of reactions needed to create small organic molecules, such as drug compounds. The pathways that the tool suggests look just as good on paper as those devised by human chemists.

The tool, described in Nature on 28 March, is not the first software to wield artificial intelligence (AI) instead of human skill and intuition. Yet chemists hail the development as a milestone, saying that it could speed up the process of drug discovery and make organic chemistry more efficient.

On Tuesday, Waymo announced they’d purchase 20,000 sporty, electric self-driving vehicles from Jaguar for the company’s forthcoming ride-hailing service.

Waymo, Google’s sister company within Alphabet, held a press conference in New York for the unveiling of the vehicle, and most of the stories revolved around the luxury SUV’s look and feel.

But the company embedded a much more significant milestone inside this supposed announcement about a fancy car. With orders now in for more than 20,000 of these vehicles and thousands of minivans that Chrysler announced earlier this year, Waymo will be capable of doing vast numbers of trips per day. They estimate that the Jaguar fleet alone will be capable of doing a million trips each day in 2020.

How private data sources, when leveraged responsibly and collaboratively, can provide insights for addressing the challenges and opportunities of migration.

Speaking to a standing-room only audience of students and professors, Bengio, who has had some 300 articles published, and whose work has been cited over 100,000 times, discussed the neural architecture behind adversarial networks, and how such networks’ core dichotomy — discriminator versus generator — makes it possible for the latter to “learn” how to create data sets by seeing what the discriminator regards as real and fake.

Stanford, CA Tuesday, April 24, starting at 5 p.m., Center for Advanced Study in the Behavioral Sciences
at Stanford University. “A conversation featuring two 2017-18 CASBS fellows – Stanford professor Nate Persily, an expert on law, democracy, and the internet; and Carrie Cihak, a senior policy expert and practitioner at one of the most innovative county governments in the U.S.” [free, registration required]

NVVL (NVIDIA Video Loader) is a library to load random sequences of video frames from compressed video files to facilitate machine learning training.

Many of the leading approaches in language modeling introduce novel, complex and specialized architectures. We take existing state-of-the-art word level language models based on LSTMs and QRNNs and extend them to both larger vocabularies as well as character-level granularity. When properly tuned, LSTMs and QRNNs achieve state-of-the-art results on character-level (Penn Treebank, enwik8) and word-level (WikiText-103) datasets, respectively. Results are obtained in only 12 hours (WikiText-103) to 2 days (enwik8) using a single modern GPU.

Matchbox enables deep learning researchers to write PyTorch code at the level of individual examples, then run it efficiently on minibatches.