SammoJ

Are these numbers for real, and are these all Bay Area? I'm further North and pull no where near these numbers as an ML PhD with a lot of experience now in the tech industry. Have I goofed on all negotiations?

I didn't look too hard but didn't see any similar free services with full DNS management and an API for cache validation. Would love a suggestion here too!

Seeing a link to the dailymail is a hint that HN may be jumping the shark. This is a UK rag of a paper.

Clearly this driver is outside of the recommended highway driving scenario. There are multiple warnings both audible and big red indications on the sensor dash that automation is about to stop.

Would we have a problem if I set cruise control in my Mazda to 60MPH then drove off a tight bend? No, we would call this driver error.

I was also desperate to see this. And rightly so, here's what looks like a prototype link from above - and it's beautiful!

https://www.youtube.com/watch?v=-drmRYsvUaM

You can really appreciate the difference between the gentle movement induced with mechanically actuated magnets, and the nasty instantaneous movement which would happen with electro.

Please show a paper where fine-grained vehicle classification in unconstrained images is anywhere near this performance from 20 years ago. You will not be able to, because it wasn't.

This is brilliant, thank you.

On the Musk transcript I found this formatting confusing: "Have you played Kerbal Space Program?

What do you think SpaceX uses for testing software?"

I can't access Reddit to see what the original comment was.

I don't see anything new here, but for any practioners like myself Scikit-learn and Spyder were the Python tools which finally moved me from a die hard MATLAB junkie.

I grabbed the handy Anaconda package from here: https://store.continuum.io/cshop/anaconda/

Within about two weeks, and with a little bit of discipline, I became a MATLAB->Python convert. Spyder is a solid IDE, and Anaconda comes with all the packages you need (i.e. Matplotlib...).

All that's left now is to find a solution to MATLAB's excellent debugging. You can break to pydb in Spyder but the debug environment is nowhere near as functional as iPython.

So I had always heard this is the case, and now believe this must be a common misconception. See Table 6 in this file from the Office of National Statistics:

http://www.ons.gov.uk/ons/publications/re-reference-tables.h...

While from 2004 to 2012 total UK fertility increased from 1.80 to 1.98, the fertility rate for non-UK born women actually /decreased/ from 2.50 to 2.29 (albeit in a slightly messy non-monotonic fashion) while the fertility rate for UK born mothers /increased/ from 1.69 to 1.90 quite monotonically.

It would be interesting to see the stats including 2nd and 3rd generation immigrant mothers (i.e. born in UK both with parents born elsewhere).

Relevant, very similar paper input/output wise, from our resident karpathy with a detailed discussion in comments: https://news.ycombinator.com/item?id=8621658

If you read his e-mail he admits he was "completely wrong"

http://news.microsoft.com/2014/10/09/satya-nadella-email-to-...

This would work for 2D games, but 3D projection doesn't allow this. Furthermore in a 3D game there are more than two degrees of freedom.

Works great in IE11 on an X1 Carbon.

Were they playing anything or just walking around in the demo? I have one and have tested it out with many people. When people get sick it's usually because they've been in one of the less interactive demos just sort of looking around wiggling the mouse etc. The people who I let loose straight onto Half-Life 2 or other games where there is some sort of goal rarely get the same type of nausea. I think this is worth taking into account when people say "it's nausea inducing".

For insta-nausea put someone in the Tuscany demo and move them around while they wear the rift. Hilarity ensues!

Didn't hunt for the source of this one but I did find some true 4K examples here:

http://media.xiph.org/video/derf/

CTRL-F 2160p

I had to drop the www http://visamapper.com/

Which attributes would you sample from a normal distribution here? I don't see any numerical attributes where this would make sense. One could add weight, height, age etc. and sample from the relevant geographical/gender distributions?

Wasn't obvious to me on a first skim through either, but on second viewing all simple variable names are single letters - exactly as one would get through minifying. See for example here and below: https://gist.github.com/anonymous/5133829#file-simcityui-js-...

If you grab the resources you can see the lightmaps.

In a mission to gather data then clearly extra megapixels give you extra data. You cannot argue against that! We aren't talking amateur photography here where the quality of the shot is important and a decent lens beats higher megapixels, we are talking acquiring measurements of the amount of photons in a discrete spatial region. Clearly a higher resolution of sensor might allow scientists to see something not visible in lower resolution images. Although I do admit that in this case I understand the choice due to specification and testing, but to suggest that extra megapixels do not give more information is silly.

The laser is 6.5KM wide when it reaches the moon. http://en.wikipedia.org/wiki/Lunar_Laser_Ranging_experiment

Adding complementary features such as edges (I'm guessing this is what you mean by lines) tends to improve the accuracy. It would also be possible to do this in real time.

I don't have time to watch the full video so I don't know what features he is currently using, but in object detection nowadays most people are using some variant of the SIFT descriptor. These are built not upon edges but on the image gradient per-pixel. The current 'hot' feature in terms of frequency of use is probably Histograms of Oriented Gradients (HOG) which do exactly what they say on the tin: Take a region of the image and count how many times a particular gradient direction occurs and the total magnitudes. Slightly more difficult to run in real-time but libraries exist.

Same for me on FF 3.6.13 Windows7 64

The deformable part based model is pretty much /the/ state-of-the-art object detection model. There is absolutely no way it should be attributed to "Liu Liu". It's creators are Felzenszwalb et al.[1].

The performance of DPMs far surpasses any other basic model such as a Viola Jones syle Haar cascade. Also you should find with DPMs that you are able to detect faces with far higher degrees of deformation - i.e. 3/4 profile faces, rotated faces etc. DPMs are far far more robust than haar cascades but with greater computational expense. Check out the papers at [1].

[1] http://www.cs.brown.edu/~pff/latent/

Pedantically a 400% increase would be going from 1 malware app to 5. "The amount of malware is 400% what it used to be" would be 1 to 4.

I have also used libsvm a lot and can heartily recommend it - but only for non-linear kernels. If you wish to use a linear SVM (which if you aren't familiar with machine learning you should probably try first) then for your own sake try libocas:

http://cmp.felk.cvut.cz/~xfrancv/ocas/html/

It uses SVM light format and also has a mex wrapper (MATLAB). More importantly I found that for linear SVMs it was around 100-1000 times faster than libsvm (I shit ye not).

It could only render for a single viewpoint. I would imagine the simplest approach (i.e. without tracking) would be to render for the viewpoint of the closest detected head.

Translating MATLAB -> C isn't too bad and could be done without any domain knowledge. Just remember the pesky 1 start for MATLAB arrays... Having a quick glance through the code I haven't seen any major reliance on MATLAB built-ins. Also "most of the tracking code is in MATLAB" is false. The tracking builds upon the Lucas-Kanade tracker in lk.cpp which uses OpenCVs implementation. I'd estimate a couple of full days of work tops to re-write in C.

Qualifications: I have had to speed up (i.e. rewrite in C) a lot of complex MATLAB code.

You will need at least the image processing toolbox and statistics toolbox for MATLAB to run the code.

A major selling point for me is not only organising papers but the ability to annotate papers with full text search for your notes - something available with the mendeley app on the iPad. If no other reader includes this ability soon I know a lot of academics planning on getting an iPad v1 once v2 is launched.

They already have the faces annotated via the tagging system. To be really sneaky they could inject faces which have been detected but NOT tagged. However they probably have enough data anyway - enough being the largest facial recognition training data set in the world.