@@tadashimori Perhaps a stupid suggestion, but I assume you also have the book Origami design secrets by R. Lang? You could try to implement some of his rules into the learning algorithm. I'm almost tempted to work on it myself! Anyway, great stuff! And I do think you're the first one to employ AI in such a way!
@@EntropicalNature Yeah, I've read the book. I know there are lots of things to implement that would help the AI find the patterns faster, but it could also end up guiding it "too much" and become just a deterministic formula. The fitness function could be made up by direct functions, like "sum the alternating angles of each point and say if it's close to 180 degrees" or "check how many points has an even number of lines", but by doing something too deterministic you end up missing some patterns that the AI could find out. As the examples in the video, I never passed any "symmetry" parameter, but I guess it noticed that when I divide the origami in sections, most of the times two sections has the same amount of points (thus achieving symmetry) Not sure if I'm the first one to employ AI in this way, but it was fun to do this "Neural network" as fitness function. The results were kinda surprising.
Okay, I thought about it a little, and the problem with a connection matrix is that for an intermediate origami we already have over 200 points. Because the neural network needs a standardized input, I would make at least a 300 by 300 matrix, and that would require it to analyze 90000 connections… Okay, I was about to say it would be a lot, but with a few convolutions it might work really well! You’d transform a cp into a 300x300 “pixels” image and apply existing techniques to is. This is genius!
TADASHI you need to recreate this with how far ai has come in a short time, im certain you'll get wonderful results, maybe even be able to create modular origami...
Good start! I guess you used some machine learning approach (since you extracted useful features to improve the algorithm). What was the input? An image? Or a tree-based structure that relates points, lines and so on? Did you apply deep learning (CNN, GANs)? I'm very interested! I also have some experience in machine and deep learning, maybe we can collaborate!
I am a computer science student with an interest in AI. I used to really like origami and might try to make a generative crease pattern AI in the future. Maybe you input a text-based prompt and it outputs a crease pattern. The possibilities with generated origami have really just started.
Como um estudante de CIência da Computação que é apaixonado por IA, e também como um amante de origamis eu achei isso aqui SENSACIONAL Adoraria ver mais sobre essa IA
As a suggestion you could take a look into generating CPs by using a GAN and specify the decision criteria :) Also I am really interested in these things and the possibly to combine this with origami would be a dream come true, let me know if I can help you with anything, would love to be a part of this!
Good idea! Never thought about something like this. You could try to force it to include bird or frog base 🤔. Maybe then results would be even better. Although I guess that technically it wouldn't be fully created by AI.
We can direct it to learn some bases. The same way I “filtered” the information before, but the only problem would be that it would add some bias to using these bases and make it lose some creativity
@@tadashimori Nice, but what is the architecture of the fitness function? Also, have you tried using deep reinforcement learning to do the same thing? It is somstimes faster to train.
@@sac666_ That's the thing that I find interesting tbh. I don't know what the neural network considers more important, so I can't really talk much about the fitness function. I will try other techniques for sure, this one was actually quite crude, and the program is a mess because I kept changing the program over the previous failed attempt. I'm quite sure that after giving this first idea, a lot of people will be able to come up with even better ideas really quickly.
I will after I finish correcting the code for these small problems, like making a symmetrical cp. Also, because I developed the code while having new ideas, the code is really ugly at the moment, with several unused functions.