AI Model Training and Accuracy

  • 2
  • Question
  • Updated 2 months ago
This question has a couple of pieces to it:

AI Model Training:
All AI models require some level of training or confirmation.  For example, if I have an algorithm designed to tell the difference between cat and dog pictures, I'd feed it a ton of pictures marked correctly as "cat" or "dog" ahead of time.  Then I would start feeding in pictures of cats or dogs that I didn't originally use in order to see if the model marks them correctly as "cat" or "dog".  When the AI mis-labels a dog as a cat, I'd need someone to tell the model it got it wrong.

My question is, given this is how AI model training usually works, how did you and do you continue to train your image detection models in a morally acceptable way?  Obviously the images are somewhat pixelated but, that doesn't mean that what goes on in the images isn't obvious.  If you are using a human to train the models, that would be a bad moral position.  However, if you only use a machine it could easily miss things that are porn or mis-label items that aren't.

If you don't have a way of confirming accurate or inaccurate image identification, you don't have a good model.  However, if you use people to confirm this, isn't that walking into strange moral territory?

I am fully OK with you getting as technical as you want in answering this question!  The more geeky, the better! You can start with a simple explanation and then get more complex.

How large were your initial datasets?  Where did you get them?  How often do you iterate on the model?  These are the kind of details I really like diving into!  Note: Obviously don't share the level of training to the point of giving away your complete proprietary formula but, the more information you can give, the better.

AI Model Accuracy Question:

I've had a couple of occasions where I thought the AI should have triggered something in my report and it didn't.  In one instance, there was a desktop wallpaper app on Steam that had a community screenshot that was clearly pornographic and others that were at least overtly sexual, yet none of them triggered a report.  Last night there was a nude image of Adam and Eve in a video I was watching.  Granted, this is art, but, I would have expected at least a false-positive in that regard.  Note: I'm not looking to intentionally "trigger" things and will not be looking for the images I saw on Steam again.  The other image I have less concerns about.

Given my first question regarding the training of the model, how do you ensure accuracy of the model and what are the actual triggers?  Note: I'm not looking for this information to "get around" detection but because I want to make sure that my family and I are properly protected/alerted.  No, an AI model can't capture anything but, what assurance do I have that the model is picking up/flagging the appropriate content?  Does the model take into consideration more extreme or less common types of pornography such as cartoon/anime styled porn?  How does the model tell the difference between art/acceptable nudity and flat out pornography?
Photo of ChrisDash83

ChrisDash83

  • 3 Posts
  • 2 Reply Likes
  • concerned

Posted 2 months ago

  • 2
Photo of Robert B

Robert B, Official Rep

  • 499 Posts
  • 47 Reply Likes
ChrisDash83,

Thank you for expressing interest in our product. I do want to address your questions, and I hope you’ll understand why my answers will be rather basic explanations. I mean no disrespect to the thoughtfulness you displayed in putting the post together.

We’ve been in this field for nearly 20 years and have a healthy company culture. In the course of our existence, we have taken great care to create tools and systems that allow employees to engage in tough work in the safest ways possible. That care was applied in the context of our original service (Internet Accountability) and it continues in the context of our new service (Screen AccountabilityTM). Over the course of time, and changing technology, our commitment to having a healthy team has not changed.

Regarding the technical details of AI training and accuracy: those details are proprietary and while we appreciate transparency as much as anyone else, we also need to protect our business interests. I will say that our accuracy exceeds AWS Image Recognition. But, while that’s a good level of accuracy, we are looking to make that even better.

Our current AI collects images about once per minute. We have seen cases where people quickly see an image (like the Adam and Eve moment you mentioned) and it doesn’t show up in the report. In such cases, it’s possible that the image was displayed so briefly that it wasn’t captured by the AI. As you can imagine, such a scenario fuels our on-going development of the AI.

Thanks again for asking about our product. I do hope you understand my reasons for being basic with my answers. If you are interested in joining our team, please check us out on Glassdoor.com.

Robert
Photo of ChrisDash83

ChrisDash83

  • 3 Posts
  • 2 Reply Likes
Hi Robert,

Thank you so much for your reply.  I fully understand the necessity to protect the proprietary nature of the AI engine itself.  I also didn't mean to question your level of integrity to your employees.  In fact, the concern you reflect in your reply demonstrates the kind of culture Covenant Eyes has within its walls.

Honestly, I never personally considered joining the Covenant Eyes team.  Right now it doesn't appear that you have any positions that directly fit my skillset.  I have a long history in customer service, training, and also have a passion for finding solutions to larger issues in workplace environments.  That includes helping to develop ideas for new processes that improve both internal and external customer experience, along with a little project management.  In short, I am a customer experience chameleon who adapts to fit many needs within a business.

That being said, I will be happy to submit my resume to your company for consideration.  I'd only ask that if you think there is anything which might fit my skillset that you please share that with the hiring manager.  I personally am using your product both for myself and my family.  I know your company's reputation for a quality product, which is why I switched from a different solution to you.

My resume will be submitted with the email address I use for this account on it.

Thank you for your response!

God Bless,
Chris
Photo of ChrisDash83

ChrisDash83

  • 3 Posts
  • 2 Reply Likes
Hi Robert,

I uploaded my resume and a cover letter to your job board.

Thanks,
Chris