Physicists think they found the rule behind modern AI

AI systems now combine many types of data, including text, images, audio, and video. But building these systems is still complicated.

Developers must choose the right algorithms for each task, and that process can be slow and uncertain in the fast-moving field of multimodal AI.

Researchers at Emory University say a new mathematical framework could make this easier.

Published in The Journal of Machine Learning Research, the framework organises AI methods around a simple idea: compress large amounts of data while keeping only the information that helps make accurate predictions.

Many modern AI systems already follow this principle.

The researchers formalised it by grouping AI techniques based on how their loss functions keep or discard information.

A loss function measures how far a model’s prediction is from the correct answer. During training, the model adjusts itself to reduce this error.

The team calls their system the Variational Multivariate Information Bottleneck Framework.

Instead of designing a new loss function for every problem, developers can adjust the framework to decide which information a model should keep and which can be removed.

Early tests show the approach can help researchers design new algorithms, estimate the amount of training data needed, and identify potential weaknesses earlier.

Because the method removes unnecessary information, it could also reduce computing demands and improve efficiency.

The researchers believe the framework may help guide the development of more accurate and efficient AI systems, and they hope others will use it to design models for complex scientific problems.

Someone looked at the chaos of AI models and said: “we need a periodic table for this mess.” - MG