How to build a working AI only using synthetic data in just 5 minutes | Hacker News Reader

How to build a working AI only using synthetic data in just 5 minutes

danrose.ai

79 pointsjetbiscuits29 comments

It should not take you 5 minutes to make an image classifier in 2022. 30 seconds is a more reasonable amount of time. Dalle is trained using CLIP which you can just use as a zero shot classifier directly, no need to waste time generating images or training a model at all. Just type in the names or descriptions of your classes and your done! Way easier than this :)

I’ve heard such claims for a long time, I can likewise create a classifier out of a simple dice role. It doesn’t say anything about how good it is.

Most software applications have low tolerance for error rates, the ones that do have big money being spent on ensuring their accuracy is better than everybody else’s.

So while you can make a classifier out of anything in Y time, that doesn’t say anything about whether it’s of any practical use.

No, CLIP is indeed that good. The robustness of its embeddings is the entire reason why VQGAN+CLIP works and can stablely generate images close to the text prompt.

I'm sure it's better than a dice role, but does it beat a modern classifier trained on the domain specific data?

EDIT: I raise this issue as over-promises are the death nell for software. Overpromising capability leads to disappointment.

See the zero-shot performance of CLIP: https://openai.com/blog/clip/

It's definitely better performance than what you'd get working in 5 minutes from more conventional approaches.

It tends to, yes. I suggest reading the paper as they discuss this very thing in detail.

Normally, this type of comment is Hacker News reductiveness, but yes, image classification via CLIP is that easy, especially with Hugging Face's API for it: https://huggingface.co/docs/transformers/model_doc/clip

I created a Python package to generate image embeddings from CLIP's vision model without requiring a ML framework (https://github.com/minimaxir/imgbeddings ), and a simple linear classifier on those embeddings does the trick, demo here: https://github.com/minimaxir/imgbeddings/blob/main/examples/...

How big of a model is CLIP? If you're building a phone app that classifies dogs, you may not want to require that it runs some multi-billion parameter monstrosity to perform its comparatively simple task. There is lots of value in building a compact model. "Just" typing in the names ignores the compute you need to have behind the scenes.

A compact model is a constraint that changes the problem entirely and doesn't discredit the quick-but-effective approach that works for nearly every other use case.

There are various sizes of CLIP, many are not enormous. For example, one of the base models is just a standard resnet50. So very usable on a mobile device.

Are you saying that DALL-E is impressive? On fediverse it's used for jokes and memes, because it's really, uh, ugly? simplistic? using obvious components in each image. To my eye, it looks like trickery. Maybe the "full, paid, commercial" model and outputs are better; i'm not sure.

I'm actually looking for a decent classifier / object recognition platform to sort on the order of millions of images coarsely - as it stands all of the ones i've tried can't determine if an image is drawn/painted or a photograph, for instance, which reduces my enthusiasm of the whole field.

On the other hand, audio AI/ML stuff - such as spleeter - impresses me, as i can't do that stuff by hand.

The original DALL-E was never released. This is a smaller model made by volunteers.

Did you consider using CLIP like parent comment said?

Dude youve seen dalle 2?

I think a more appropriate title would be: How to click an upload data button on a website.

I’m curious why you would need to generate the images using another AI. There are masses of free high quality 3d models out there and even if you can’t find the model you need you could always use photogrammetry to create one from a real world example. After you have the subject you can render it in a game engine like Unreal Engine from as many different perspectives and lighting conditions as you want. Since the engine is programmable you could even automate this part to a large degree and adding other confounding objects or backgrounds would be simple as well.

What's really interesting with 3d models is that you could give the classifier motor control within it's environment i.e. it could "step to the left" and see how the banana changes in it's visual field. This would allow the classifier to integrate it's movement with the visual stimulus and build a far richer model. It also gives the classifier access to an active learning strategy: predict how the appearance of the banana changes when it "steps to the left", try it, then evaluate the difference and refine.

Interesting post, but it seems fragile enough that I'm not sure it could work if you try to classify a yellow apple.

Have you seen Roboflow Universe? They have thousands of projects from their community already labeled for download and use. I always start there when I want t to start a new model. Recently for my trash picking up robot I found a good litter dataset to start from. https://universe.roboflow.com/

Title: "How to build a working AI (...)".

Text: How to build an apple-or-banana image classifier.

Ah. I see. It's an allegory for the state of modern machine learning research.

First define what is a "working AI" even mean Be careful AI is not an iPhone app you will not find any funding for BS in this env.

>In a short while, it has gone from being an experimental technology. To something, I would hesitate to use for production AI solutions.

What?

I've worked in the AI space, but mostly on the backend tweaking algorithms so I guess I'm asking this as (mostly) a laymen: Isn't using AI to generate training data to train another AI fraught with peril?

And… I am still waiting with baited breath for something like this that would work for tabular data.

Training images today is equivalent of launching a web server using one line of code

Don't waste your time. There are so many statistical atrocities in this article that it makes me shiver.

Where? I don't see any statistics in the article.

Probably meant that using synthetic data is recycling within a closed system.

I guess that's possible, but it's not clear to me that that matters. It'll still work right? What do we care?

Well if you know the world only from the words of Jack knowledge will be limited, right? And if you fed it back to Jack to have it tell you again the same in a different way, than you have a loopy closed system with definite risks of non enhancement.

Reply on news.ycombinator.com