Skip to content
 
If you can read this, either the style sheet didn't load or you have an older browser that doesn't support style sheets. Try clearing your browser cache and refreshing the page.

(Unite.ai)   Since AI companies are too cheap to develop their own image datasets, the old ones they're constantly using could open up a hilarious new sector in internet security   (unite.ai) divider line
    More: Interesting  
•       •       •

973 clicks; posted to STEM » on 29 Nov 2021 at 2:50 PM (7 weeks ago)   |   Favorite    |   share:  Share on Twitter share via Email Share on Facebook



22 Comments     (+0 »)
View Voting Results: Smartest and Funniest
 
2021-11-29 1:50:51 PM  
Fark user imageView Full Size
 
2021-11-29 2:41:55 PM  
Obligatory:

BLIT

First we use graffiti to kill Skynet, and then...
 
2021-11-29 3:46:53 PM  

croesius: [Fark user image 585x832]


I too enjoy ping-pong ball omelettes, what's your point?
 
2021-11-29 4:43:32 PM  
I was showing Google Maps street view to my kids, and they noticed how everyone's faces were blurred.  I told them it was because the AI just knows how to recognize faces and blurs them to protect people's privacy.  It happens that we were looking at the campus where I work, and in some of the shots, you can see my boss walking along.  For whatever reason, the AI chose to blur his ass.  We lol'd.
 
2021-11-29 4:56:26 PM  

Ambitwistor: croesius: [Fark user image 585x832]

I too enjoy ping-pong ball omelettes, what's your point?


That poor computer had never been so sure of anything in its life.
 
2021-11-29 6:30:37 PM  
I know a thing or two about images of cheap Als

i.redd.itView Full Size
 
2021-11-29 6:42:40 PM  
Name 1 useful image AI application that is really useful and not could be useful.
 
2021-11-29 7:55:39 PM  

mr0x: Name 1 useful image AI application that is really useful and not could be useful.


The one that takes pictures of  your license plate and mails you a bill so you don't have to stop at a tollbooth.
 
2021-11-29 11:21:53 PM  

RogermcAllen: mr0x: Name 1 useful image AI application that is really useful and not could be useful.

The one that takes pictures of  your license plate and mails you a bill so you don't have to stop at a tollbooth.


Not AI.
 
2021-11-30 4:50:03 AM  

mr0x: Name 1 useful image AI application that is really useful and not could be useful.


I have AI-generated captions for all my 50k+ pictures I took during the years, which helps me find pics taken when playing tennis, or have mountains in them, or my motorcycle, etc. I believe Google does this automatically too, I just used Microsoft's service through an API.

You can also train it to recognise specific people, so you get all the pics of grandpa taken over the years.

At work we've trained an object detection model to recognise specific stamps on old digitized (scanned) documents which helps to classify them correctly in the digital archive. This would be a massive pain to do manually as we've got 70 years worth of documents.
 
2021-11-30 4:59:01 AM  

mr0x: RogermcAllen: mr0x: Name 1 useful image AI application that is really useful and not could be useful.

The one that takes pictures of  your license plate and mails you a bill so you don't have to stop at a tollbooth.

Not AI.


OCR is a form of AI, in fact most modern OCR systems use some type of convolutional neural network in its detection process.
 
2021-11-30 6:21:37 AM  

neaorin: mr0x: RogermcAllen: mr0x: Name 1 useful image AI application that is really useful and not could be useful.

The one that takes pictures of  your license plate and mails you a bill so you don't have to stop at a tollbooth.

Not AI.

OCR is a form of AI, in fact most modern OCR systems use some type of convolutional neural network in its detection process.


Actually, no. They may use a fourier transform, but that isn't a neural network. It's a math algorithm that produces the same result every time. Depending on what information you mask, you highlight different information.

Now these techniques were originally inspired by neural networks. Basically mathematicians and engineers figured out what the neural network was actually doing. (Though, to be fair, most of AI development is actually throwing out ideas that don't work.)

But what is the Fourier Transform? A visual introduction.
Youtube spUNpyF58BY
 
2021-11-30 6:30:00 AM  
(And by "inspired by" I mean the algorithms were developed in the 19th century and sat on a shelf until the computer engineers got tired of banging their head against a wall.)
 
2021-11-30 6:33:56 AM  
Fourier Optics used for Optical Pattern Recognition
Youtube Y9FZ4igNxNA


And if you are interested, here is how to recognize basic letters with a modified form of holography. You can accomplish the Fourier transform with light interference.
 
2021-11-30 7:11:47 AM  

Evil Twin Skippy: neaorin: mr0x: RogermcAllen: mr0x: Name 1 useful image AI application that is really useful and not could be useful.

The one that takes pictures of  your license plate and mails you a bill so you don't have to stop at a tollbooth.

Not AI.

OCR is a form of AI, in fact most modern OCR systems use some type of convolutional neural network in its detection process.

Actually, no. They may use a fourier transform, but that isn't a neural network. It's a math algorithm that produces the same result every time. Depending on what information you mask, you highlight different information.



I believe most products out there have moved on to ML-based approaches, they just generalize better and deal with noise much more robustly than traditional FFT (if you're training correctly and not overfitting, of course)

ABBY FineReader is a popular product which has been using ML, including deep learning, for a while now.

Microsoft's OCR service is now deep learning based.

Google engineers have also given talks about their vision cloud service and it's also convnet-based for most tasks including OCR.
 
2021-11-30 7:14:08 AM  

neaorin: Evil Twin Skippy: neaorin: mr0x: RogermcAllen: mr0x: Name 1 useful image AI application that is really useful and not could be useful.

The one that takes pictures of  your license plate and mails you a bill so you don't have to stop at a tollbooth.

Not AI.

OCR is a form of AI, in fact most modern OCR systems use some type of convolutional neural network in its detection process.

Actually, no. They may use a fourier transform, but that isn't a neural network. It's a math algorithm that produces the same result every time. Depending on what information you mask, you highlight different information.


I believe most products out there have moved on to ML-based approaches, they just generalize better and deal with noise much more robustly than traditional FFT (if you're training correctly and not overfitting, of course)

ABBY FineReader is a popular product which has been using ML, including deep learning, for a while now.

Microsoft's OCR service is now deep learning based.

Google engineers have also given talks about their vision cloud service and it's also convnet-based for most tasks including OCR.


You believe or you know?

And full disclosure I write expert systems for a living.
 
2021-11-30 7:17:38 AM  
(And nobody uses "machine learning" in production. The might use a weight mask derived from machine learning. But you don't sic a learning network onto uncurated data. Unless you are an idiot or an automated car company.)
 
2021-11-30 7:25:53 AM  
Evil Twin Skippy:
You believe or you know?

And full disclosure I write expert systems for a living.


I believe. Don't know for sure that everyone does it. I provided examples of big companies doing it, obviously I can't find examples for every company out there.

Also full disclosure, I do Computer Vision using deep neural nets for a living. Not for OCR specifically, but enough to understand why they are more robust if you train and validate correctly. Especially when packaged in a product or service which is likely to encounter extremely varied and noisy data in the hands of customers.
 
2021-11-30 8:02:28 AM  

neaorin: Evil Twin Skippy:
You believe or you know?

And full disclosure I write expert systems for a living.

I believe. Don't know for sure that everyone does it. I provided examples of big companies doing it, obviously I can't find examples for every company out there.

Also full disclosure, I do Computer Vision using deep neural nets for a living. Not for OCR specifically, but enough to understand why they are more robust if you train and validate correctly. Especially when packaged in a product or service which is likely to encounter extremely varied and noisy data in the hands of customers.


Apologies.

I get people in chats who can't set the clock on their dashboard who tell me how computers work. Hard to tell an expert from a crackpot.

I'm kind of surprised you don't use some of the analytic techniques to clean up you inputs before feeding them into a [whatever neural nets are going my theses days.]

But then again in the days of ASICS, float operations were cheap, it was memory that was expensive.
 
2021-11-30 8:15:56 AM  
(Fun fact for the non software engineers in the audience. MPEG and Jpeg compression are both applications of the discrete cosine transform. The compressor fits blocks of the image to a set of frequencies that approximate the signal. Then it stores that frequency data in a compact format. When you open the image/video your computer fills all of the data by inverting the transform. This is, needless to say, a pretty numerically intense process. Thus why your smart phone needs more power than a 1980s mainframe.)
 
2021-11-30 8:06:22 PM  

Evil Twin Skippy: (And nobody uses "machine learning" in production. The might use a weight mask derived from machine learning. But you don't sic a learning network onto uncurated data. Unless you are an idiot or an automated car company.)


Imagining the end result of that is hilarious, though...
 
2021-11-30 8:09:57 PM  

It'sMorphin'Time: Evil Twin Skippy: (And nobody uses "machine learning" in production. The might use a weight mask derived from machine learning. But you don't sic a learning network onto uncurated data. Unless you are an idiot or an automated car company.)

Imagining the end result of that is hilarious, though...


It puts the laughter back in manslaughter for sure
 
Displayed 22 of 22 comments

View Voting Results: Smartest and Funniest

This thread is closed to new comments.

Continue Farking




On Twitter


  1. Links are submitted by members of the Fark community.

  2. When community members submit a link, they also write a custom headline for the story.

  3. Other Farkers comment on the links. This is the number of comments. Click here to read them.

  4. Click here to submit a link.