MIT made AI design less messy
Matt V
April 24, 2025
AI SCIENCE
MIT researchers have put together something pretty cool: a periodic table of machine learning algorithms. Okay, we think it’s cool.
It maps out how over 20 classical methods are connected, offering a clearer way to mix ideas and build better models.
At the heart of it all is one equation.
This equation explains how different algorithms learn relationships between data points, even if they go about it differently.
By using it as a foundation, the team created a new framework called Information Contrastive Learning (I-Con), which shows how familiar tools are more alike than we might think.
It’s not just for show.
The table actually points to "missing" algorithms, ones that should exist but haven’t been discovered yet.
By filling one of those gaps, the team built an image-classification algorithm that beat a top-performing model by 8%. Isn’t science amazing?
The idea came about by chance when MIT’s Shaden Alshammari noticed two separate algorithms shared the same maths.
Here’s what you should know:
MIT built a periodic table linking 20+ machine learning algorithms through one shared equation.
It highlights gaps where new algorithms could be created, and one already outperforms current models by 8%.
Researchers now have a shortcut to experiment and design smarter AI methods.
Same brain, different outfits
That kicked off a bigger project to group more methods under one roof.
Now, instead of starting from zero, researchers can use the table to mix and match ideas more easily.
The research will be shared at the International Conference on Learning Representations and was developed with input from Google AI and Microsoft.
A real throwback to my science class.