How can I find abstract features among drawings?

AI Discussions
,
1

Hello everyone,

I am a little new to machine learning, and only know very basic architectures like the usual ANNs or other basic algorithms like k-Means.

Let’s say I have drawings from people, and I want to find common patterns or “motives” among them. How could I achieve this, if there even is any feasible way?

(They are (very) abstract drawings, so no guarantee that there are perfect drawings of a house or a tree on them, haha. So, it really is about finding recurring abstract patterns, themes, …)

I asked ChatGPT about it, and got a few recommendations:

  • using a pretrained model like ResNet or something more adapted for stylized stuff, and simply seeing what it finds

  • using a CNN, but cutting off the last (classification) layer to get access to the features it “found”

  • using an Autoencoder to see what (latent?) features it learned

All of these would conclude with dimensionality reduction if needed, and clustering via k-Means, according to ChatGPT!

Now, I dont know much about such advanced architectures and wanted to ask, if those methods are feasible, or would be the simplest option!

I would gladly appreciate any help or advice on how to approach this, what things to look into, or honest comments if this is not really feasible!

Also, just ask if any more information is needed!

Thank you!

2

It is feasible but you need some dataset to train the model, it maybe be better to use a pretrained model and do transfer learning because you dont have a big dataset.

Wether you remove final layers or not should be a decision you make after trial and test. In Tensorflow Advanced Techniques Specialization they go through transfer learning. What Chatgpt also suggests seems to make sense too and maybe more complete, make sure to follow up with asking further question with chatgpt and how to implement it, also ask for sample code its possible that it can provide to you.