Scientists Have Created a Periodic Table for AI

“Hearst Magazines and Yahoo may earn commission or revenue on some items through these links.”

MIT researchers found that different algorithms can all be grouped into a ‘periodic table’ of AI.
The idea for the table was an accident that emerged from identifying similarities between two algorithms.
Having a periodic table of algorithms will make it easier to figure out which ones to use together depending on the desired properties.

In high school chemistry, we all learned that the elements of the periodic table are arranged according to their properties (and most of us will never forget having to fill them in on at least one test). Now, computer science is taking advantage of the usefulness of the periodic table to help them achieve their own scientific ends.

MIT researchers have created a periodic table that orders AI algorithms based on the properties they share with a mathematical framework known as information contrastive learning (I-Con). The table describes how algorithms find the data points where they are linked—a.k.a. where they have that mathematical DNA in common—and internally approximates those links. An individual algorithm can maximize efficiency by minimizing the differences between approximated connections with other algorithms and connections those algorithms are trained to make.

Just like heavy metals and noble gases, algorithms on this AI periodic table are grouped according to what they have in common—both with I-Con and with each other (I-Con is a loss function, which basically gauges a model’s performance by figuring out how much it deviates from correct predictions). The table itself was something of an accident that emerged from commonalities between classical machine-learning algorithms. When Shaden Alshamarri, an MIT graduate student, was studying the clustering algorithm—which can learn to sort images into clusters depending on how similar they are—she found that it reminded her of a different algorithm known as contrastive learning.

“I-Con reveals that many seemingly disparate methods including clustering, spectral graph theory, contrastive learning, dimensionality reduction, and supervised classification are all cases of the same underlying loss function,” she and her team said in a study that was recently presented at the 2025 International Conference of Learning Representations (ICLR).

The contrastive learning algorithm was inspired by human brain function in the visual cortex—parts of the ventral visual cortex are involved with recognizing and categorizing objects, and a model trained in contrastive learning can learn to accurately carry out visual tasks by contrasting objects against each other. Alshamarri discovered that a mathematical function (I-Con) could show how the clustering algorithm she was studying was connected to this contrastive learning algorithms.

As Alshamarri and her team found relationships between the I-Con function and more types of algorithms, they realized they could arrange them in a way that mirrored the periodic table. These algorithms include dimensionality reduction (which deals with an overwhelming amount of data by reducing the features in a dataset) and supervised classification (which has the model predicting outcomes).

Just like the actual periodic table, which includes blank spaces for elements that have yet to be discovered, there are empty spaces in the AI periodic table that have been left blank for algorithms whose datapoint connections to I-Con are still unknown.

“We believe that the results presented in this work represent just a fraction of the methods that are potentially unify-able with I-Con,” Alshamarri said in the same study. “And we hope the community can use this viewpoint to improve collaboration and analysis across algorithms and machine learning disciplines.”

You Might Also Like