Can ChatGPT be used to identify a tartan?

[jsrnephdoc]

I think you will need to make "kiltGPT" and train it with the knowledge of Peter MacDonald

Taking advice from a true fountain of wisdom (Obi Wan Kenobi), this morning I remembered his "there IS no TRY, there is only DO."

So, I DID just visit the ChatGPT Website and showed its LLM the JPEG I'd snapped from a video while spinning on the exercise bike yesterday.

Less than a five seconds later, the bot responded with the correct answer: "Loud Macleod, or "MacLeod of Lewis"

I was SO amazed I decided to confirm the bot's intuition by checking the Scottish Registry of Tartans. The input "MacLeod" returned MANY choices, but none of them what I'd seen on the singer. However, when I input "MacLeod of Lewis" it revealed what I'd seen, but also besmirched my sense of good taste, noting that it was one of the "from thin air" creations of the Sobieski Stewarts!

[Jock Scot]

Taking advice from a true fountain of wisdom (Obi Wan Kenobi), this morning I remembered his "there IS no TRY, there is only DO."

So, I DID just visit the ChatGPT Website and showed its LLM the JPEG I'd snapped from a video while spinning on the exercise bike yesterday.

Less than a five seconds later, the bot responded with the correct answer: "Loud Macleod, or "MacLeod of Lewis"

I was SO amazed I decided to confirm the bot's intuition by checking the Scottish Registry of Tartans. The input "MacLeod" returned MANY choices, but none of them what I'd seen on the singer. However, when I input "MacLeod of Lewis" it revealed what I'd seen, but also besmirched my sense of good taste, noting that it was one of the "from thin air" creations of the Sobieski Stewarts!

Whatever the history of the MacLeod of Lewis tartan may be, many MacLeods of Harris used to wear the MacLeod of Lewis as their "dress" tartan. I am not sure if they still do, as I don't do formal events these days. But my Grand Father and Father used to wear it on very formal occasions. I on the other hand usually wore the MacLeod of Harris for formal events as well as everyday attire and I understand, that many make that choice these days.

[figheadair]

Taking advice from a true fountain of wisdom (Obi Wan Kenobi), this morning I remembered his "there IS no TRY, there is only DO."

So, I DID just visit the ChatGPT Website and showed its LLM the JPEG I'd snapped from a video while spinning on the exercise bike yesterday.

Less than a five seconds later, the bot responded with the correct answer: "Loud Macleod, or "MacLeod of Lewis"

I was SO amazed I decided to confirm the bot's intuition by checking the Scottish Registry of Tartans. The input "MacLeod" returned MANY choices, but none of them what I'd seen on the singer. However, when I input "MacLeod of Lewis" it revealed what I'd seen, but also besmirched my sense of good taste, noting that it was one of the "from thin air" creations of the Sobieski Stewarts!

If you're saying that ChatGTP identified this as MacLeod then it's wrong and the diagnosis is a fine example of the maxim Rubbish In-Rubbish Out.

The tartan is Ayrshire Gold, a 2023 fashion design by Clyde Kilts, an Ayrshire based retailer.



As if one wasn't bad enough, they also do an Ayrshire Silver.

[bookish]

AI can only reproduce what it's trained on, and how well the training is managed so that it produces statistically relevant responses. I doubt a general or generative AI can produce positive ID of a tartan. I would also be careful about what training data you supply a tartan AI. The real meat of tartan is the sett threadcount and color definition.

I don't know if the tartan register creates pixel-perfect renderings of tartan for their website or not. Certainly not with the small thumbail images. You can get images from vendors and weavers. I wouldn't necessarily trust exact threadcounts from those images, either. So, quality training data needs to be secured.

My 2 cents.

[jsrnephdoc]

AI can only reproduce what it's trained on, and how well the training is managed so that it produces statistically relevant responses. I doubt a general or generative AI can produce positive ID of a tartan. I would also be careful about what training data you supply a tartan AI. The real meat of tartan is the sett threadcount and color definition.

Akin to the "abilty" of a phone to recognize recorded music, where the recognition is actually of a digital and inaudible code into the music that the "listening" device can recognize.

Thanks for that.

[bookish]

Akin to the "abilty" of a phone to recognize recorded music, where the recognition is actually of a digital and inaudible code into the music that the "listening" device can recognize.

Thanks for that.

I wanted to highlight that there are different tools within the AI sphere. General and Generative are their own sets within that sphere. I'm a novice, so of course, check me on that with reliable sources that aren't selling you on the "magic" of AI. I am also biased against AI for philosophical and economic reasons, so, do consider my bias as well. I discourage people using AI for simple answers to questions. I encourage people to use AI to create assistance in learning and internalizing how to learn to answer questions themselves. I'll step off my soapbox, here. I'll also note that I am not trying to criticize you for testing out Chat GPTs capabilities, here, and that my bias may come off as criticism. You don't know what something can do until you've tried it.

If tartan identifying was something you were interested in from a software product, I'd think that there are machine learning models out there that will get a software engineer up and running and training on sets of images. I just wouldn't trust General and Generative AI agents like Chat GPT on something as specific and niche as tartan.

[jsrnephdoc]

I General and Generative are their own sets within that sphere.

I didn't know that they were not one and the same. So, if I were looking for a Generative AI answer to my question, I'd be making the following assumptions:

1. The generative AI database would recognize that the needed dataset would include numerical data such as thread counts and sett size
2. Those would be prioritiezed over inferences about color, unless a dataset that specified color based on pantone numbers or something equivalent existed.
3. only THEN could a generative AI engine answer my question, and even then it's more complicated, because the "look" of a tartan, including sett size depends on other factors as well, such as how thick the individual threads are and how tightly they're woven.

And my guess is that that sort of cataloging has not been done.

[bookish]

I didn't know that they were not one and the same. So, if I were looking for a Generative AI answer to my question, I'd be making the following assumptions:

1. The generative AI database would recognize that the needed dataset would include numerical data such as thread counts and sett size
2. Those would be prioritiezed over inferences about color, unless a dataset that specified color based on pantone numbers or something equivalent existed.
3. only THEN could a generative AI engine answer my question, and even then it's more complicated, because the "look" of a tartan, including sett size depends on other factors as well, such as how thick the individual threads are and how tightly they're woven.

And my guess is that that sort of cataloging has not been done.

You know, your points are allowing for me to get my thoughts in order a little better. I probably have given the wrong impression about thread count and color.

1. I think this approach is a non-AI way of identifying tartan. If you had a relational database of sett colors and threadcount, and an algorithm that analyzes images of tartan to approximate the size of color blocks, then you could determine if a picture of a tartan was more likely tartan x or tartan y.
2. I think it is enough to try to say, "this block of color is likely within the [general color name] color space." Though, this probably is a problem for the non-AI way of trying to identify a tartan with an algorithm.
3. AI is trained on images in different ways, but generally AI doesn't know or care what the context of an image is. It doesn't know what tartan is, and doesn't relate context to its determination. It just says, "I have been trained to say that when something kinda looks like this, it is tartan. This is more yellow, so there is a likelihood that it is Macleod." It doesn't look up "loud Macleod" in a database, it just has to be told a lot what is and what isn't Macleod, and probably has a statistical weight that says when something kinda looks like this, it is probably Macleod.

I bet someone in the mill industry is trying to train a model on tartan...

I have a friend who actually works on AI stuff, so I'll ask him about this use case.

[OC Richard]

The tartan is Ayrshire Gold, a 2023 fashion design by Clyde Kilts, an Ayrshire based retailer.



As if one wasn't bad enough, they also do an Ayrshire Silver.

Thanks Peter!

As we know some of the big firms involved in Kilt Hire trot out entire lines of tartans from time to time, and I had guessed that it might be from one of those. I didn't know about Clyde Kilts.

Witness MacGregor And MacDuff's line of "Mist" tartans https://macgregorandmacduff.co.uk/co...20Mist%20Kilts

and McCall's "Pride" tartans https://www.mccalls.co.uk/pride-pages/weathered-pride/

and Houston's "Bute" tartans https://www.kiltsforhire.com/search?q=bute

[bookish]

Since I brought up that there are different types of AI or machine learning, I didn't want to leave it at that. So that anyone can look into it themselves, I asked my friend, Dr. Shearman about this. My questions were:

1. Is AI really the right tool, or would a more deterministic set of algorithms be better?
2. What methods of training an AI on the images be best?
3. What general class of AI would this fall under?

His responses were:
1. Yep, AI or ML (machine learning) is the right tool for this task.
2. Assuming you don't have any a priori structure to your images, I would try fine tuning a vision language model.
3. Vision language models are what you want. Inputs: (image, text) -> text. Such that (img, "does your image contain a person wearing a kilt?") -> "yes"

He went on to explain that there is a way to segment the image to focus only on the tartan, and use other image processing for the identification. You can also fine tune the model at the final stages of development on targeted examples, which requires a lot of labeled data. He also stated that a small convolutional neural network might be able to solve this problem directly.

Much of this is specialized knowledge that I would have to look into more to understand (like what a "convolutional neural network" is). So, I think I'll leave it at that for folks to investigate if this is really a problem space you find interesting. I don't know if the finished product would justify the cost in time and money to develop, but there we are.