ways for machine learning systems to break

The NIPS paper out of Google lists them all:

  • Boundary erosion
  • Data dependencies
  • Feedback loops
  • Machine learning-systems anti-patterns
  • Configuration debt
  • Changes in the world outside of the ML system

Very accessibly written and easy to follow if you have any experience putting modeling systems to work in any capacity.

Full disclosure: I work at Google, though not on the engineering side of the house.

Bayesian Program Learning makes it to the New York Times

Via the NYT:

Researchers at the Massachusetts Institute of Technology,  and the University of Toronto reported a new type of “one shot” machine learning on Thursday in the journal Science, in which a computer vision program outperformed a group of humans in identifying handwritten characters based on a single example.
The program is capable of quickly learning the characters in a range of languages and generalizing from what it has learned. The authors suggest this capability is similar to the way humans learn and understand concepts.

The researchers apply a new technique known as Bayesian Program Learning. There's a nice discussion of the topic on Quora, of which one of the commenters says you can think of BPL as "deep learning without a massive training set." The paper referenced by the article is Human-level concept learning through probabilistic program induction by Lake et al. and is freely accessible via Science.

andrew ng on deep learning at baidu

Baidu is an incredibly nimble company. Stuff just moves, decisions get made incredibly quickly. There’s a willingness to try things out to see if they work. I think that’s why Baidu, as far as I can tell, has shipped more deep-learning products than any other company, including things at the heart of our business model. Our advertising today is powered by deep learning.

Read up on his full interview with WSJ. This is, of course, the same Andrew Ng of Stanford machine learning fame.

contributor by google: pay not to see ads

For all of the complaints about ads being a bad way of supporting the "free" web, it's the ad technology companies that are innovating.

The Mountain View, Calif.-based tech giant is testing a program called Contributor by Google with 10 publishers, including The Onion, Imgur and Mashable, in which users can pay $1, $2 or $3 a month to not see ads. A "thank you" message will appear in place of the promos, and the user will pitch in pennies from their $1 to $3 allotment every time they visit one of the sites.

More at Adweek. And here is the official site.

FnordMetric ChartSQL: Charting using SQL

FnordMetric ChartSQL allows you to write SQL queries that return charts instead of tables. The charts are rendered as SVG vector graphics and can easily be embedded into any website and customized with css in order to build beautiful dashboards

Lots of examples here, all of them with some degree of interactivity. The framework comes bundled with a standalone HTTP server app, and the charts are simple and effective:


More here.