Google Has Made It Simple For Anyone To Tap Into Its Image Recognition AI (gizmodo.com)

An anonymous reader quotes a report from Gizmodo: Google released a new AI tool on Wednesday designed to let anyone train its machine learning systems on a photo dataset of their choosing. The software is called Cloud AutoML Vision. In an accompanying blog post, the chief scientist of Google's Cloud AI division explains how the software can help users without machine learning backgrounds harness artificial intelligence. All hype aside, training the AI does appear to be surprisingly simple. First, you'll need a ton of tagged images. The minimum is 20, but the software supports up to 10,000. Using a meteorologist as an example for their promotional video was an apt choice by Google -- not many people have thousands of tagged HD images bundled together and ready to upload. A lot of image recognition is about identifying patterns. Once Google's AI thinks it has a good understanding of what links together the images you've uploaded, it can be used to look for that pattern in new uploads, spitting out a number for how well it thinks the new images match it. So our meteorologist would eventually be able to upload images as the weather changes, identifying clouds while continuing to train and improve the software.

Comments

They want the crowd to train their machine

[Anonymous Coward]

because it's free....

Re:They want the crowd to train their machine

[TheRaven64]

I wonder if that's actually true. On the one hand, they get a lot of training data for free. On the other hand, how well are the protected against adversarial training? If a botnet signs up for a million accounts, submits 20 photos with correct descriptions for 19 and incorrect ones for the remaining one, is it possible for the attacker to teach the Google algorithm some nonsense things? What if I'm more subtle and upload 100 pictures of dogs and tag them as dogs, but the pictures are carefully crafted to share more characteristics with other pictures of cars than of dogs? Can I train it to recognise all cars as dogs (last year researchers were able to alter a picture of a dog or a car by one pixel and have this system recognise it as the other). A system like this is only as good as its training set - remember that this is the same system that decided that all black people were gorillas because the only black face in its original training set of photos from Google employees was a gorilla at a zoo.

Re:They want the crowd to train their machine

[140Mandak262Jamuna]

Adversarial training is not limited to Artificial Intelligence. Natural Stupidity also suffers from the same fundamental flaw.

How else can you explain political parties training their flock to vote against their own self interest?

Re:They want the crowd to train their machine

[swillden]

Nope, you are testing and helping their A.I. to learn using your data, at your expense, and THEY (Google) get all the benefits.

Even if they get some benefit, it's silly to say that they get all the benefits. I mean, if you do this it's because you'd like to have an AI trained to so some image classification job, and it's easier for you to use Google's service than to build and train your own. So it must be fulfilling some need of yours, and doing it more cheaply/easily than you could do it yourself. That's value.

Re:They want the crowd to train their machine

[deodiaus2]

When I started out as a programmer, you never wanted to be anyone's beta tester. Now, people give up their free time to do so!

Re:They want the crowd to train their machine

[TheConway]

Free to Google, which is what he probably meant.

The thirty fourth rule.

[queazocotal]

'Using a meteorologist as an example for their promotional video was an apt choice by Google -- not many people have thousands of tagged HD images bundled together and ready to upload.'

I suspect quite a few of us have tagged HD images bundled together and ready to upload.

Re:The thirty fourth rule.

[PolygamousRanchKid]

I suspect quite a few of us have tagged HD images bundled together and ready to upload.

My list of of twenty pictures:

Hillary.
Donald.
A can of Cheez Whiz.
A fractal cow: http://mndl.hu/2008-02-01-frac...
A Bitcoin.
A Jai Alai Cesta.
John Small Berries.
A build break.
An iPhone battery.
Bacon.
Bigfoot.
Queen Elizabeth's Crown Jewels.
An ingrown toenail.
Twenty years to life, with no chance of parole.
"The Economist" international Big Mac index.
A Nobel Peace Prize.
An Ig Nobel.
Winter Storm f "Friederike".
A Ford F-450 Super Duty Limited.
Slashdot.

Ok, Google AI . . . get at it . . .

Re:The thirty fourth rule.

[Nutria]

Mine pics are of 20 different gorillas.

double take

[Froze]

Did anyone else read that title where the second capital T was an F?

Is this actually AI?

[Known Nutter]

Serious question... is this really Artificial Intelligence? Or is this really Self-Adapting Algorithm? For that matter, is "Self-Adapting Algorithm" what "Artificial Intelligence" actually is?

"AI" is such a hyped up, overused term that I just can't tell what's what anymore. I'm old.

Re:Is this actually AI?

[Anonymous Coward]

"AI" is whatever we don't know how to get a computer to do.

Once we know how to get a computer to do it, we give it a different label, and it becomes "no longer AI".

Re:For now.

[sTERNKERN]

No, it is not free. Just as their search engine is not free (where you pay with your own personal data) this time it is your time, effort, data you provide.

Re:For now.

[asylumx]

I believe he meant it's free for *Google* -- they won't have to pay people to feed it data.

Cloud OpenML as an Encrypted Communication Tool

[PeteJanda]

Admittedly, this concern may be a bridge too far even for the tinfoil hat crowd. But...

If I were a bad guy, knew that intelligence agencies have compromised electronics down to the firmware and hardware levels and needed to securely communicate with other bad guys, then I'd develop image + label data to train Google's service to spit out plaintext results from certain image sets. My compatriots would run images of dogs, cats, etc. through OpenML and receive labels like "Bomb" "Building" "Corner" "Columbus" "Central Park".

Good luck to the good guys when trying to pick up such e-mail traffic.

Re:Cloud OpenML as an Encrypted Communication Tool

[CustomSolvers2]

Although theoretically possible, the scenario you propose seems too complex to ever become a reality. It would be much easier to create a picture-based encryption system; in fact, I developed one of those some time ago.

Re:Cloud OpenML as an Encrypted Communication Tool

[PeteJanda]

A picture-based system is only as strong as the protection of its cipher, which can be compromised in its storage medium (i.e., somewhere, somehow there's a key saying "dog" equals "bomb"). With OpenML, which is meant to be a turnkey solution for mitigating complexity, the cipher is the hidden layers. The interesting aspect here is that the state of the art in AI can't fully explain how those hidden layers arrive at their probabilistic results. Hence a potentially strong system.

Re:Cloud OpenML as an Encrypted Communication Tool

[CustomSolvers2]

You are making a pretty good point, but practically speaking the requirements for implementing both alternatives seem still way off. Also bear in mind that images contain orders of magnitude more information than plain text, what means excellent conditions to create very strong encryptions.

The system you propose could be way harder to crack, but also to train (not just making sure that the algorithm behaves as you want, but also hiding any suspicious activity from Google). The strongest point of image-based encryption, at least when properly done, is the difficulty to even know whether there is encrypted information there at all. That's why I think that, if you are able to set up a proper image-based communication, the strength of the encryption algorithm wouldn't be one of your main concerns.

Did google ever fix the gorilla problem?

[kalieaire]

Did google ever fix the gorilla problem?

http://www.dailymail.co.uk/sciencetech/article-5270891/Google-bans-word-gorilla-racist-Photos-app.html

If you tap into the Google

[Dirk Becher]

the Google also taps into you.