It can be broadly applied across diverse imaging systems in photography, microscopy, and astronomy.

We developed a method to estimate mutual information directly from noisy measurements of unknown objects. It requires only measurements and a noise model.

Information determines what's possible: With more information, algorithms can discern different objects, and they can reconstruct better, classify more accurately, or extract otherwise imperceptible details.

Imaging systems are encoders. They map objects (scenes, cells, galaxies) to measurements through lenses, sensors, and detectors.

Imaging systems have traditionally been designed to produce pictures for human eyes. But increasingly, measurements get processed by AI.

What if we designed them to maximize information rather than visual appeal? Could we see the universe, diagnose diseases, and capture photos better?

#NeurIPS2025