Something scary: AI judged photo contest results

[sirch]

I've mentioned in other posts how AI can generate relatively sophisticated assessments of photographs, and while I don't "know" , I suspect you could fairly easily choose to train your AI bot on Ansel Adams photographs or Gary Winogrand photos etc and then use a Landscape AI bot trained on Ansel Adams pictures to assess a Landscape competition.

I suspect the difficulty is that AI might be artificial but it is in no way intelligent. As far as I am aware both Adams and Winogrand only produced monochrome work so an AI trained on their work would only judge B&W as "good". Also you need a large corpus of images with the full spectrum from good to bad and some way for the AI to learn what is good and what is bad.

The thing no one talks about in "AI" is the transfer function (or activation function) which is the mathematical function that calculates the output of a neuron from its input. These functions are programmed into the AI by a human and hence why AI in its current form is in no way intelligent.

 

[ancient_mariner]

These are being judged by an "AI judge" in real time with a rolling top twenty being displayed. First prize €500, and all you need to do is upload your files.

When I first looked, I thought they were nearly all awful, On a second look, the pictures have changed, and to my eyes, there are several that are OK, but the majority....

AestheticsCompetition

competition.excire.com

Excire who are behind this competition use AI to keyword photographs.

The blurb that goes with the competition says:

"The juror: our AI

You don't show your photographs to a large jury, but only to a single juror: Excire's artificial intelligence.

No personal preferences or tendencies limit its judgment.

Excire's artificial intelligence objectively evaluates your photographs.

To do this, it has learned from hundreds of thousands of photos what distinguishes good photos from bad ones.

This sample data was curated by numerous professional and ambitious amateur photographers and thus perfectly reflects aspects of image quality and aesthetics."

Is the site still functioning for you? I've tried on a couple of different browsers (one without my usual security measures) and also without the tracking data in your link, but there are no images being shown.

 

[Nod]

Have to say that I prefer photos that are on the rare side rather than the horrendously overcooked ones that appear to be A"I"'s choice at the moment!

 

[AndrewFlannigan]

These functions are programmed into the AI by a human and hence why AI in its current form is in no way intelligent.

Agreed.

I've had conversations with a psychologist who specialised in neural development. He argued that no machine system could be genuinely intelligent unless it was self booting, like humans. Perhaps it's not surprising that he had read Robert Heinlein's novel, "The Moon is a Harsh Mistress", which was written in the 1960s. One of the two major characters is a computer, which "becomes intelligent" rather than being programmed to appear intelligent.

My aquaintance said that he thought Heinlein, albeit writing long before any of the tools existed to to test the theory, had hit on the key difference between intelligence and an appearance of intelligence. In the psychologist's opinion, real intelligence depends on constant learning about both objects and their context plus the creation of self awareness, which he postulated as a multitude of feedback loops that were constantly increasing, as we experience stimuli.

I'm pretty sure that this is a rabbit hole, down which we could lose ourselves for hours!

 

[myotis]

I suspect the difficulty is that AI might be artificial but it is in no way intelligent. As far as I am aware both Adams and Winogrand only produced monochrome work so an AI trained on their work would only judge B&W as "good". Also you need a large corpus of images with the full spectrum from good to bad and some way for the AI to learn what is good and what is bad.

Well, that of course is the issue, AI isn't all that intelligent, which is part of the reason I'm opposed to using to judge photography competitions,

But, I was just suggesting that if you use it within certain limits it could be more useful than trying to make a model that tried to encompass all types of photographs.

Hence the suggested very limited scope ie one trained on Ansel Adams pictures assuming the judges were only interested in a winner that was an "Ansel Adams" type picture. And make use of some appropriate rule based decision making.

I posted a link earlier comparing two approaches to using AI as a Capture One assistant. One used a regular AI that searched the Internet for answers, and the other was expert trained. The difference between them was striking.

I'm still not keen on it's use for this at all, but it might be useful if you are sifting through a million competition entries or 5000 wedding photographs, where you want to weed out those that stand no chance of winning/being presented to the bride. There are already AI programs that do the latter.

As an aside Adams had at least one book in colour and did commercial work in colour. He also worked with Kodak on their first digital camera (which I assume was colour), something he was very excited about .

The thing no one talks about in "AI" is the transfer function (or activation function) which is the mathematical function that calculates the output of a neuron from its input. These functions are programmed into the AI by a human and hence why AI in its current form is in no way intelligent.

I have no idea about calculating the output of a neuron, but does anyone actually believe AI is "intelligent". The stuff, I've seen is pretty impressive, while at the same time seriously flawed, including some very stupid mistakes. But the speed it can do things, would seem to make up for the rather mixed results it produces, as long as it's subject to human scrutiny

I confess to not knowing much about AI, but as I mentioned earlier I've built some Bayesian Probability models, and the outputs from the model is controlled by the quality of the data used to build it, how we choose to interpret that data, and the rules we end up creating. So lots of human intervention.

 

[myotis]

Is the site still functioning for you? I've tried on a couple of different browsers (one without my usual security measures) and also without the tracking data in your link, but there are no images being shown.

I've been using Chrome, and it's still working for me.

I've now also pasted the link into Safari and it's working as well.

The pictures are taking about a minute to load, but as I said my connection is very slow.

I assume others have been successful.

I can't think of what to suggest

 

[ancient_mariner]

I've been using Chrome, and it's still working for me.

I've now also pasted the link into Safari and it's working as well.

The pictures are taking about a minute to load, but as I said my connection is very slow.

I assume others have been successful.

I can't think of what to suggest

I wouldn't worry. Waited over 2 mins, different browser again, nothing doing.

FWIW there used to be a site called (IIRC) Lightbox to which one could upload photos. The images would arrive at the front of a stream, and could be voted on. Images receiving votes/likes were kept near the front of the stream, with those that had fewer recent likes dropping back to eventual obscurity. Pictures were required to be relatively small, and the only way to get attention was to make an image striking, bright and contrasty with plenty of colour and a simple, strong subject. Anything subtle, complex or finely detailed would quickly disappear.