Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Sergarr]

There's actually no "differentiation factor" feature extraction going on here:

2.1 ARCHITECTURE
During training, the input to our ConvNets is a fixed-size 224 × 224 RGB image. The only preprocessing
we do is subtracting the mean RGB value, computed on the training set, from each pixel.
The image is passed through a stack of convolutional (conv.) layers, where we use filters with a very
small receptive field: 3 × 3 (which is the smallest size to capture the notion of left/right, up/down,
center).

3.1 TRAINING
The ConvNet training procedure generally follows Krizhevsky et al. (2012) (except for sampling
the input crops from multi-scale training images, as explained later). Namely, the training is carried
out by optimising the multinomial logistic regression objective using mini-batch gradient descent
(based on back-propagation (LeCun et al., 1989)) with momentum. The batch size was set to 256,
momentum to 0.9. The training was regularised by weight decay (the L2 penalty multiplier set to
5 · 10−4
) and dropout regularisation for the first two fully-connected layers (dropout ratio set to 0.5).
The learning rate was initially set to 10−2
, and then decreased by a factor of 10 when the validation
set accuracy stopped improving. In total, the learning rate was decreased 3 times, and the learning
was stopped after 370K iterations (74 epochs).

3.2 TESTING
At test time, given a trained ConvNet and an input image, it is classified in the following way. First,
it is isotropically rescaled to a pre-defined smallest image side, denoted as Q (we also refer to it
as the test scale). We note that Q is not necessarily equal to the training scale S (as we will show
in Sect. 4, using several values of Q for each S leads to improved performance). Then, the network
is applied densely over the rescaled test image in a way similar to (Sermanet et al., 2014). Namely,
the fully-connected layers are first converted to convolutional layers (the first FC layer to a 7 × 7
conv. layer, the last two FC layers to 1 × 1 conv. layers). The resulting fully-convolutional net is
then applied to the whole (uncropped) image. The result is a class score map with the number of
channels equal to the number of classes, and a variable spatial resolution, dependent on the input
image size. Finally, to obtain a fixed-size vector of class scores for the image, the class score map is
spatially averaged (sum-pooled). We also augment the test set by horizontal flipping of the images;
the soft-max class posteriors of the original and flipped images are averaged to obtain the final scores
for the image.

Essentially, this algorithm classifies things as similar, according to how close they are to each other after projecting a huge dimensional point corresponding to an image in the space of all possible images of a given size, to a smaller (but still very large) dimensional space.

Also, note that this is essentially equivalent to a young childs level of understanding of things, because of how very little data is fed to these networks. I expect the results to start improving vastly, once we manage to actually figure out a good way to automate gathering, combining and re-using all available data. Right now, every time we switch to a different neural network (or any other learning method), all active information stored in it is essentially lost completely, which means the next one has to start from scratch again.

Can you imagine how a human society would look, if every single human had his memory wiped at least every six months or even more often, and was mostly kept in isolation from other humans? I'd imagine there wouldn't be much understanding among humans on anything in that case, either.

[Sergarr]

I mean, it identifies whether a group of images is similar to another. It doesn't have any fundamental basis of what's actually in each image. There's a reason why in your presented image, the perspective and angle of the shot is almost identical between the selections. There's no "understanding" of food, just a matching algorithm between sufficiently similar rgb distributions.

They don't look all that similar in color distributions to me. You also have to account for the fact that there are a lot of kinds of food with very similar color distributions, and this algorithm has successfully managed to differentiate from them.

I mean, that's not even mentioning the little fact that the article, on the basis of which this neural network was done, directly says that it uses image data practically pixel-by-pixel, with no "RGB distribution" calculations or anything.

[Sergarr]

By that very wide definition of "rgb distribution", human vision (from a single eye) is also "just a matching algorithm between sufficiently similar rgb distributions".

I also don't quite understand how do you define and quantify "understanding" in this case. You are expecting an algorithm that only has data about fixed, static images, to somehow "know" the functionality of objects and their logical relationships between each other, both of which are inherently dynamic properties and impossible to know from static images. These are sure some unreasonable expectations. Do you really expect a human, who has only been ever exposed in their life to static images of the outside world, to have any understanding on relative dynamics of objects on these images?

[Shadowlord]

By that very wide definition of "rgb distribution", human vision (from a single eye) is also "just a matching algorithm between sufficiently similar rgb distributions".

No? Natural light is not composed of combinations of a single wavelength each of red green and blue light, it's all photons with different wavelengths which basically just looks the same to us because reasons.

[i2amroy]

Two pages late on my response to this, but here it goes anyways.

Can we, though? Last I checked even at the peak of our destructive capability -- which we're not really at anymore -- we couldn't really manage that. Destroy human civilization, sure, wreck huge chunks of the biosphere, definitely, but destroy anything on the surface? Is shit out there that takes more than a sustained nuclear bombardment to get rid of. Honestly, our best bet is probably what we're doing environment wise, and even the worst of that is unlikely to surpass previous mass extinction events.

We're pretty damn good, but we're not that far, yet. We've still got a ways to go before we really can even really top a sustained wide-scale algae bloom, and when you're being outperformed by pond scum, you don't get to claim superiority

I did some calculations a handful of months back and came to the conclusion that at our height (and still today) we have the cumulative destructive capability to basically destroy anything on the earth's surface. However it's important to recognize that while we can destroy anything, what we cannot do is destroy everything. Even at our height of nuclear destruction we would have been hard pressed to simply cover 20-30% of the earth's land area (funnily enough about the same as a very rough estimate of how much of the land area we actually use) in non-overlapping nuclear devastation, and nowdays we can probably only hit somewhere between 10-15% of the total land area, based on the accuracy of various estimates about current stockpiles.

(It's also important to realize that in the case of trying to destroy certain things [such as say, mountain ranges], we would be much better off planting our bombs inside of the mountains like we were blasting rather than trying to devastate it with a top down bombardment.

[Furtuka]

http://futurism.com/amputee-gamer-gets-a-prosthetic-bionic-arm-with-light-charger-and-drone/

[Bumber]

http://futurism.com/amputee-gamer-gets-a-prosthetic-bionic-arm-with-light-charger-and-drone/

No rocket punch upgrade yet?

[Shadowlord]

http://www.nytimes.com/2016/05/26/health/alzheimers-disease-infection.html

Or if you hit a paywall, http://www.msn.com/en-us/news/us/could-alzheimer%E2%80%99s-stem-from-infections-it-makes-sense-experts-say/ar-BBtutfg?ocid=ansmsnnews11

[Frumple]

Wunnit they saying something about something fungal related not too long ago? Looks like they were, checking google, though those links are talking about something else. Wonder if it's going to end up that there's actually several different causes causing the same symptoms... know they have some kind of name for that, for all I've completely forgotten it.

[Shadowlord]

This mentions that fungal infections could do it too - if they're right, it's the innate immune system trapping invaders in amyloid cages, whether they're bacteria, fungi, whatever.

[Frumple]

Ah, no... was talking about this. Not sure how much it's been followed up since, but th'study mentioned apparently found the same type of fungal infection in the brains of every alzheimer's effected cadaver they checked, within a fairly small sample group.

[Shadowlord]

http://www.wsj.com/articles/a-first-look-at-americas-supergun-1464359194

Railguns have for years been limited to laboratories and videogames.

Former President Ronald Reagan ’s Strategic Defense Initiative—the so-called Star Wars missile defense—at one time envisioned using the railgun to shoot down nuclear missiles. Those plans were stalled by 1980s technology. One problem was that the gun barrel and electromagnetic rails had to be replaced after a single shot.

The Navy now believes it has a design that soon will be able to fire 10 times a minute through a barrel capable of lasting 1,000 rounds.

I guess we're in the future now.

[Loud Whispers]

No, we're in Future R&D hell

[Spehss _]

Posting to watch science.

[Starver]

Posting to watch science.

Do not just watch...  Observe!