148 karma · joined May 26, 2012
It was great how he spent a lot of time on logistic regression before delving into SVM's or Neural Nets - it was much easier to understand the cost functions & regularization for other types of classifiers after having understood those for logistic regression.
My takeaway: if you can avoid adding risk to your systems by using more complicated models, you should.
Can you elaborate on what kind of policies towards entrepreneurship are better in Alabama and Arkansas? Seems to me that SF Bay Area and Los Angeles are both flourishing entrepreneurial havens [1]. I'd argue that the culture of innovation in California outweighs any economic policies that may be adverse to entrepreneurs.
[1] http://techcrunch.com/2012/11/20/startup-genome-ranks-the-wo...
> Implementing these taxes nation-wide would ensure the U.S. has the same air quality as Los Angeles.
California still has among the worst Air Quality ratings in the US [1]. Isn't it worth calling out that California needs to improve in this area? For instance, better public transit?
> California has the highest paid teachers in the U.S. and by focusing on teacher pay rather than student achievement (students rank 48th in math and 49th in reading), they have the best teachers in the country.
This hardly seems like a reason for California to pat itself on the back. If we are ranked so low in math & reading, what is the metric we're using to evaluate our teachers and determine that they're the "best in the country"?
[1] http://www.stateoftheair.org/2012/states/california/los-ange...
It really is hard to separate ML & statistics - any competent practicioner of ML appreciates the statistical achievements that made Machine Learning methods possible. And statisticians must understand that to help automate decision-making systems, using learning methods/boosting is a viable option.
The debate around nomenclature (ML/stats/AI) seems limited to the academic community. Most data scientists I've met tend to accumulate a repertoire of tools from different fields, rather that side with either Machine Learning or Statistical communities.
For a while, I favored "The Worst" - I liked having a few missing keys on my laptop, and the challenge of programming with less RAM than I should have. I liked having an old car that barely scraped by. The freedom of not feeling attached to my belongings made me feel a sense of purity - like my code on github is the only possession I really cherish.
After a few car breakdowns and failed hard drives, however, I've learned to pick a few things worth focusing a little effort on!
Is anyone clear on whether the raid on his house happened in Belize, or how he got back there between the alleged murder time and now?
[1] http://www.oas.org/juridico/mla/en/traites/en_traites-ext-us...
[2] http://www.usnews.com/usnews/news/articles/050815/15belize_3...
A good friend of mine just flew down to Santiago for Startup Chile. He's definitely enjoying it, but before he left he was concerned that getting out of the US entrepreneurship scene might make it harder to reach US consumers and get relationships with investors. On the flipside, S.C. seems focused on creating an international community of entrepreneurs - I'm told there's teams from 33 countries.
It'll be interesting to see how many of these teams will end up in YC.
This brings up bigger questions with regards to authentication - how does the 'api' service know that a request is coming from a correctly authenticated user? Since on heroku, the 'api' service would be externally exposed, I'm trying to decide if there's a good way to correctly authenticate/authorize in such a way that the 'api' service understands.
Thoughts?
1) Try to be more selective with the meetups you attend. Particularly, meetups focused around a programming language or a specific technology (Python/Hadoop/Node.js/etc) are likely to have more dedicated technical attendees. Bonus points if the meetup doesn't allow vendors/recruiters - that generally means that they like to keep the conversations focused on technology, rather than sales or job offers.
2) Hackathons. It's often encouraged to show up without an idea, form a team, and work on an idea for 24-48 hours. This is a great way to test-drive a partnership with someone who you might invite to work on a side-project. Chances are, by the end of the hackathon, there's mutual respect/trust among team members; the bonds you form by hacking together for a day or two are going to be far more conducive to working on new projects than the bond between you and your twitter followers.
3) Open-source projects. Pick an open-source project that you find particularly interesting. Sign up for the developer and user mailing lists, ask questions and answer them where you can. If you ask the developers where a good place to start helping out is, they'll often give you some easy ways to help out (improving logging / documentation / code style is often an easy first patch).
Hope this helps!
Copenhagen has done studies [1] indicating that every mile driven accounts for some net economic loss (due to the harmful health effects of sedentary lifestyles, pollution, the cost to the commuter of gas/cars, the cost to society of maintaining automotive infrastructure), where very mile ridden constitutes a net economic gain (decreased healthcare costs due to being more active, lower personal cost of owning a bicycle, etc).
If the US spent more time looking at commuting from an economic and health perspective, we'd probably find that more government action is needed, beyond "Bike to work day"
I use R to prototype machine learning techniques on a small set of data, or visualize some summary statistics. But if I want to run K-Means Clustering or Support Vector Machine algorithms on 1,000,000,000 rows of data, I've found that running R on Hadoop is tricky. There are some libraries out there ( for example, RHadoop https://github.com/RevolutionAnalytics/RHadoop/wiki/rmr ) but they require writing your algorithm in such a manner that algorithms must be adapted to run within map() and reduce() functions. My understanding is that the built-in functions that make R so useful will often not adapt well to a mapreduce algorithm.
From what I've seen, once an algorithm is prototyped in something like R/Matlab, if the data size warrants it, it's best to re-write the algorithm in Java MapReduce or use Apache Mahout.