Sneak Peak: Mona Lisa Drawn using 512 Circles

Published on 2024-10-18.

This is my final semester at the University of Toronto where I’m currently taking a Neural Networks and Deep Learning course taught by Prof. Collin Raffel. Taking the course has led me to, once again, revisit an old project of mine to apply my learnings.

The goal of the project is to make a program that can draw any target image using only circles. Of course one can do that with a greedy algorithm and a large number of circles (even as extreme as having one circle per pixel!). However, in my opinion, the “better” solution would try to replicate the image with as few circles as needed to ultimately produce novel and artistic results.

The old solution, used a hill-climbing algorithm to randomly change circle parameters over hundreds of thousands of iterations. The new approach uses gradient descent through backpropagation, a smarter ML technique that converges after only a few hundred iterations (aka epochs).

I plan on writing a more detailed article about the new solution, but for now, I will leave you with those two pictures as an amuse-bouche:

Gradient Descent

_{Mona Lisa drawn using 512 circles after 1 thousand epochs in ~7 minutes}

Hill Climbing

_{Mona Lisa drawn using 256 circles after 1 million iterations in ~18 minutes}