Supply generates its own demand; by making captioning even cheaper, it can increase the demand for transcription services and people to check it over. There are a lot of podcasts and YT videos that could benefit from transcriptions but it's too expensive now.
We have to hardwire architectures because we can't learn architectures yet, or to put it another way, backpropagation doesn't yet work on hyperparameters as well as on parameters. Hyperparameters should be learnable as in theory there's nothing special about them (it's models all the way down!) - a hyperparameter is merely a parameter we don't yet know how to learn - and this has been demonstrated: http://jmlr.org/proceedings/papers/v37/maclaurin15.pdf "Gradient-based Hyperparameter Optimization through Reversible Learning"
"Tuning hyperparameters of learning algorithms is hard because gradients are usually unavailable. We compute exact gradients of cross-validation performance with respect to all hyperparameters by chaining derivatives backwards through the entire training procedure. These gradients allow us to optimize thousands of hyperparameters, including step-size and momentum schedules, weight initialization distributions, richly parameterized regularization schemes, and neural network architectures. We compute hyperparameter gradients by exactly reversing the dynamics of stochastic gradient descent with momentum."
But it's not feasible yet. Once it is, you can imagine collapsing the whole neural net zoo: you merely specify the input/output type/dimension and then it starts gradient-ascent over all the possible models as tweaked by internal hyperparameters.
It could also reflect what is most distinguishable. Which is easier for a NN to confidently distinguish: black pubic hair on black skin, or black pubic hair on white skin? Darker nipples on black skin, or darker nipples on white skin? etc You're doing gradient ascent on confidence of classification, not simply trying to find a plausible input, but the maximal input. There's no reason to expect this to be racially unbiased as there are simple objective reasons that higher contrast would be useful. (Similarly, I would not be surprised if a face recognition NN worked better on Europeans rather than Chinese, for the prima facie reason that they have more variable facial features and other aspects like multiple hair colors other than black.)
And since Yahoo needs to detect porn of all races and there's plenty of black porn out there, it would be odd if their porn detector had such a huge gaping hole in it.
What sort of performance can be expected compared to running in the terminal? How large NNs will this scale to in practice? I see a 50-layer resnet is mentioned; but not 1000-layers?
It's interesting to read about the debate between doing editing/screen-drawing on the physical terminal ('smart' hardware) and doing it all in software. The wheel of reincarnation in action! Something we've seen in tech constantly, the balance between local vs distant computing determined by bandwidth/latency... Lately, smartphones, desktops and VR or sound: should the processing be done in the headphone/headset or away in in the PC/smartphone?
How would you know if they have been? As the human population shoots up massively, car-miles have increased enormously over the past century, and rural areas become denser, the number of deer-car accidents would skyrocket even if they have successfully been evolving to reduce risk.
The obsession with information theory here seems like a classic nail-hammer thing. The number of bits my tests convey is totally useless to think about and certainly not worth spending pages on. All I want from my tests from a code base I maintain for thousands of patches is a tiny fraction of a bit: did my latest change break an important behavior or invariant encoded in a unit test? If I only screw up once in every 100 patches, then formally, my unit tests are doing all that work to emit 0.01 bits of information (-log(99/100)), which is formally a totally irrelevant thing to know about my unit testing framework. ('Hey Joe, what have you been up to?' 'Fixing my unit testing framework - I'm up to 0.03 bits per patch!' 'I see.')
This is perhaps the stupidest and most hysterical thing I've ever read on Stross's blog. I don't think I need to explain why Brexit is not going to lead to the https://en.wikipedia.org/wiki/European_Potato_Failure and a million deaths from starvation.
You can definitely see where this is going. My first thought was the same: 'we need eyetracking for foveated rendering anyway, so we can get realistic eyes for free', and if you can do that, you can track the eyebrows and muscles around the eye (doesn't need great fidelity), and I wonder if that gets you all the way to the rest of the face as well? Can you smile/frown without it tugging on the parts closer to the eyes which the future headsets can observe?
"Experts said it was likely that the NSA or FBI had approached other Internet companies with the same demand, since they evidently did not know what email accounts were being used by the target. The NSA usually makes requests for domestic surveillance through the FBI, so it is hard to know which agency is seeking the information.
Reuters was unable to confirm whether the 2015 demand went to other companies, or if any complied.
Alphabet Inc's Google and Microsoft Corp, two major U.S. email service providers, did not respond to requests for comment."
It's not a hard question to answer. You either are or are not searching all emails in realtime at the behest of the NSA.
> If you think the Internet is as safe and controlled as a shopping mall, you probably should be reading Krebs on Security more.
That's an amusing comparison, given how much of Krebs focuses on offline ATM skimming, copying credit cards at point-of-sale terminals, hacking major retailers's CC databases, and using stolen cards at retail and mall stores to cash them out...
You could probably apply the same code. The dataset ("acquire 75,471 sketches of 12,500 objects") sounds adequate, and if not, can be boosted by first training a CNN to do photo->sketch (throwing away information is usually easier than imagining it) and using that to boost the dataset for sketch->photo.