genericpseudo
genericpseudo on Hacker News
I like "Machine Learning: A Probabilistic Perspective" by Kevin Murphy more, but this really is a matter of taste; both are excellent.
Take as much probability and linear algebra as you can conveniently do – as much for the intuition as for the symbol-manipulation mechanics – and don't underrate the importance of domain expertise in any problem you get interested in!
The Shipping Forecast is a very significant cultural reference in the UK. It crops up all over the place. Britain is _fundamentally_ an island and a seafaring nation, and that's something Americans miss; you're never more than seventy miles from the sea. It's as iconic as, I don't know, Thanksgiving football in the US; it's a thing everyone knows about without explanation.
This is from an album which sold over 1.2m copies in the UK; one of the biggest records of the 90s: https://www.youtube.com/watch?v=SD8gO8TAr4s
Density functional theory, for example, is taught at undergraduate level as part of both chemistry and physics triposes. Probably materials science too, and it certainly used to be an option within earth sciences (as part of the mineral physics path).
Collaborative filtering is machine learning, so if you consider AI machine learning, definitely yes.
The team behind Detectron have published an enormous amount of really good research, but the Detectron codebase struck me as "good research code" rather than something you'd ideally want in production.
Not at the kind of resolution you'd want to be using on, e.g., Twitch. In that setting, you could just use chromakey, though? That's '70s technology, cheap and very reliable.
If you're interested in this but have no background, the best place to start is "Fully Convolutional Networks for Semantic Segmentation" –
https://people.eecs.berkeley.edu/~jonlong/long_shelhamer_fcn...This is a very active field of research. Another thread worth pulling on is Mask R-CNN: https://arxiv.org/abs/1703.06870
It's not quite as simple as "this one has highest mAP, let's use it"; the tradeoffs are complex. In particular, as you can see in the image here, one thing DeepLab doesn't do is segment instances – so you get a mask of "people", not a mask per person. Mask R-CNN does a better job on that by design, because it predicts both bounding boxes and a mask per bounding box.
> Stitch Fix makes it a major brand point that computers are an important part of the process
(which even if you disregard everything else is a really good line to take if you're after recruiting good machine learning engineers!)
This means that codec development makes no sense for anyone who doesn't own either a large
distribution platform or a large
playback platform. (Much of the research has been done by middleware companies attempting to tax the two; their business goes from royalties to work-for-hire, at best, which is way less attractive for them).
This is fine – in a macroeconomic sense – but of course it sucks if you're one of the companies being disrupted.
One of four (MXNet and CNTK alongside TF and Theano), and the Amazon deep learning API forked Keras to default to MXNet support before it was really ready - which irked the Keras authors quite a bit.
Python has been the language of choice for many physicists (when not doing Fortran) since the early 2000s.
PyTorch, Theano (rip), TensorFlow itself, Numba/Blaze?
It's even more dramatic than that. Google make money on advertising, not direct sales, so what you're saying is "the (usage * advertiser desirability) from the entirety of group 2 is roughly equal to groups 1 and 3 combined".
Given group 1 is many times bigger than group 2, that's a very strong statement, but it's borne out in all the data I've ever seen, both public and private.
Absolutely, yes. Your moral argument there is pretty undeniable; I'm speaking purely to propensity-to-spend and that's a pretty narrow lens.
In fact, I'd say "target the web" if you're going for maximum accessibility and you're not driven by commercial factors, though that doesn't work for every app and the usability/discoverability issues can be real. Favoring any commercial platform as a government is a very uncomfortable place to be.
Turns out paying $700 is a pretty good proxy for "caring", though!
More "chooses to spend what money they have on a phone" plus "has the money to spend", which is almost but not quite what you're saying. I know some very rich people with whatever Android device the network gave them. They just don't care that much.
This observation is the single most important thing you need to know if you work in consumer mobile.
To first order, iPhone owners spend money. Android owners don't. This is because your average iPhone user cares more about what phone they're using and simply uses it more.
This is a first-order approximation. The small percentage of people who actively choose Android do spend and do use their phones a lot, and by goodness are they vocal, but the more useful way of thinking about the market is not two-segment, it's three-segment:
* Vast majority: don't care about their phone OS, won't pay for anything
* Significant minority: want iPhones, will most probably spend money
* Significant but even smaller minority: actively want Android, will buy premium Android phones (e.g. Nexus, high-end Samsung), will either spend money or, with roughly equal likelihood, jailbreak and pirate everything in sight.
From this perspective iOS remains the most compelling mobile OS to target. Additionally, iOS users – on average - use apps more and for longer, though again that effect is small when you control for the kind of Android devices people go out of their way to choose.
Yes, when appropriate. Think of it as a Bayesian prior...
At a previous company, I worked with someone in a recent highly-public sexism scandal (if I said who or where, I would no longer be a genericpseudo).
I knew that the person was extremely brash and was told that they indulged in "locker-room talk", though I didn't personally observe any harassing behavior. I didn't speak up about what I heard. I wish I had. I wonder, and worry, if anything as bad as what happened at their next company happened at the company I worked at. I'm really uncomfortable with the odds that it did.
The problem isn't the mods' local actions (though the mods are not nearly as assertive as I would like in this area); it's the Hacker News community's toleration of bigots and therefore the
macro action of failing to remove them.
Inaction is a form of action.
Close but not quite. The difference between (soft) SVM and a kernel linear classifier is choice of loss function; SVM minimizes hinge loss, linear regression minimizes squared loss.
(Choice of different loss functions will also give you Elastic Net, LASSO, logistic regression. From an engineering point of view I tend to think of the entire class as being different flavors of "stochastic gradient descent", in the spirit of Vowpal Wabbit etc.)
Jobs has form for this kind of thing, too (breaking something to force developer behavior
before reintroducing it when habits have changed).
Best example: the original Mac had no cursor keys (to force developers to build mouse-based UIs). Steven Levy writes about it at length in "Insanely Great".
Has Thunderbolt ports. The existing iMacs can drive external displays in addition to their internal one.
Running a data warehouse is a full-time commitment. Tons of real businesses have no software engineers.
This is a great-looking product.
That would involve
a) writing yourself an engine and that's actual work
b) that engine not being one which is used in real research
So "it depends" – for introductory teaching yes, for training researchers less so.
It looks (from the outside) like Etsy's core eng team, early, had its culture come from all of the ex-Flickr people working there.
Flickr's other direct descendant, culturally, is Slack. Those are nerdy, charming, funny cultures. "Code as Craft" is a cute pun. Taking it entirely literally is a really big point-miss.
Me of fifteen years ago would
love this for lab classes on, e.g., classical molecular dynamics.
The computational load of these simulations is, in modern terms, not large, so Emscripten-compiled simulations would be plenty fast enough, and deployment of web pages in a computer-lab environment (or for people working from their dorms!) is way easier than the alternatives.
So there are real, non-trivial applications. You wouldn't want to be running ABINIT or whatever in your web browser, but GROMACS on some small system? Not unreasonable.
Totally would :-)
They're an insurance company which monetizes through operating businesses rather than the traditional model of bancassurance.