NYU Data Science newsletter – May 19, 2015

GROUP CURATION: N/A

… SCALE: What is your role at Google? There has been some speculation about it, but not a lot of real public discussion.

Eric Brewer: I am working on stuff related to Kubernetes and containers. That’s a project I care a lot about and am certainly pushing Google to do more in that direction. That’s actually very exciting to me.

At the core of Uber’s wild success and market valuation of over $41 billion is its data and algorithmically fueled approach to matching supply and demand for cars. It’s classic economics, supposedly: “Prices go up to encourage more drivers to go online. The increase in price is proportionate to demand,” says the official Uber video explaining their surge pricing system. It’s an easy line to buy into, but is Uber’s surge pricing algorithm really doing what they claim? Do surge prices really get more cars on the road?

My analysis suggests that rather than motivating a fresh supply of drivers, surge pricing instead re-distributes drivers already on the road.

… As both the algorithms and systems for training and using neural networks have evolved, LeCun has remained active in further development, both in his role as founding director at the NYU Center for Data Science and as the director of AI Research at Facebook. He tells The Platform that while using GPUs on a single node for these workloads has been perfected, parallelizing across a number of a machines (and many thousands of GPU cores—one Nvidia Tesla K80 has over 4,800) is still a roadblock, especially for the training of neural networks, which can take several days.

LeCun and his team at the NYU Center for Data Science, are tackling this research challenge with a new eight-node Cirrascale cluster outfitted with 48 of the high-end Nvidia K80 GPU coprocessors. The goal is try to find efficient ways to maximize GPU computation across a cluster for rapid training of large neural nets.

If I hear about the recommendation algorithm one more time I’m going to lose my mind. Framing the conversation like there’s one, perfect, singular rec engine that you’ll be able to find when you come up with just the right algorithm or run just the right A/B test is missing the point. A rec engine is a navigational mechanism and you should treat it like any other tool in your design palette, like colors or page elements or site structures. There are many different recommendation algorithms and the thing that’s different about them isn’t how they maximise your favourite KPI, it’s how they treat the people that have found themselves using your product.

Videos of presentations from ICLR 2015.

Artificial Neural Networks are powerful models in Artificial Intelligence and Machine Learning because they are suitable for many scenarios. Their elementary working form is direct and simple. However, the devil is in the details, and these models are particularly in need of much empirical expertise to get tuned adequately so as to succeed in solving the problems at hand. This post intends to unravel these adaptation tricks in plain words, concisely, and with a pragmatic style. If you are a practitioner focused on the value-added aspects of your business and need to have a clear picture of the overall behaviour of neural nets, keep reading.

tl;dr: Reproducibilty is a noble cause and scientific software a promising vessel. But excess of reproducibility can be at odds with the housekeeping required for good software engineering. Code that “just works” should not be taken for granted.

This post advocates for a progressive consolidation effort of scientific code, rather than putting too high a bar on code release.

Applications are invited for an Ideas Lab on “Interdisciplinary Approaches to Biomedical Data Science Challenges” taking place from July 20 to 24, 2015 at the Statistical and Applied Mathematical Sciences Institute (SAMSI) located in the Research Triangle Park, North Carolina.

Application Deadline: Monday, May 25