#uber describing their #Bayesian inference model to predict location of a trip. Basically they build a prior consisting of favorite destinations of a particular rider, an overall location probability and overall popular locations. This is combined with a likelihood taking into account the safety of the target neighborhood, distance from the start and time of day. They achieve 74% accuracy. What I would have liked to see is a comparison to always guessing the most popular destination for a rider based on time of day and starting point - but of course that would only work for people using Uber many times.
Good news from the NIH. They have finally recognized the use of Henrietta Lacks' cells for decades, and have a new rule that those cells can only be used after approval, with two family members on the board deciding. Rebecca Skloot probably also deserves a lot of credit for highlighting the case in her book, "The Immortal Life of Henrietta Lacks".
Recent research from the Center for Talent Innovation shows U.S. women working in science, engineering, and tech fields are 45% more likely than their male peers to leave the industry within the year.
Nice progress for #earthquake research! The European Sentinel satellite system has measured interferograms of the Napa valley earthquake, meaning images of how much the earth shifted. The satellite captures before and after images and allows measuring the shifts.
The Arctic sea ice is still receding at an alarming pace! Sea ice coverage fluctuates every year, as expected the hot summers reduce sea ice, while the cold winters increase the ice levels.
The figure below shows the average 1981-2010 sea ice area (black line), while the gray shaded portion is 2 standard deviations, (95% confidence). 2012 (blue dash line) was a record low for summer sea ice, the current 2014 trend (green line) is less than the record breaking 2012 year. However, as you can see, the green line is extremely close to the bottom edge of 2 standard deviations.
Found these slides from +Max Welling and collaborators on how to speed up MCMC for #bigdata . Speed up the Metropolis step by replacing it with a t-statistics test (reduces amount of data needed to make a decision), use Langevin dynamics with Stochastic Gradient descent.