FiGuRO: Intrinsic Dimension Estimation for Multi-Modal Data

Imagine you are trying to describe a complex object, like a swirling galaxy or a bustling city, using only a few words. If you say "it's big and shiny," you've lost most of the story. If you write a thousand-page novel, you have too much detail and it's hard to find the point. Scientists call the "just right" amount of information needed to describe something its Intrinsic Dimension. Think of it as the true number of knobs you actually need to turn to recreate a shape, even if that shape is sitting inside a room with a million other things. For example, a flat sheet of paper has an intrinsic dimension of 2 (length and width), even if it's floating in a 3D room.

Now, imagine you have two different cameras filming the same event: one sees in color, the other in black-and-white. They are both watching the same "shared" action, but each camera also captures its own unique "private" details (like the color of a shirt or the grain of the film). The big challenge for computers is figuring out: How many "knobs" describe the shared action? And how many describe the unique details for each camera? If the computer gets this wrong, it might mix up the story, thinking the color of the shirt is part of the main action, or it might get overwhelmed by too much data. Getting this right is crucial for making smart AI that can understand biology, medicine, and the real world without getting confused.

Enter FiGuRO (Fidelity-Guided Rank Optimization), a new tool created by researchers to solve this exact puzzle. You can think of FiGuRO as a super-smart, self-correcting editor for data. Imagine you are trying to pack a suitcase for a trip. You start by throwing everything you own into it (too much stuff!). FiGuRO is the traveler who keeps checking the suitcase: "Is this jacket actually necessary? Can I take it out without ruining the trip?" But here's the magic: if they realize they packed too little and forgot a coat, FiGuRO puts it back in. It constantly shrinks and grows the "size" of the data's description until it finds the perfect fit.

The paper shows that FiGuRO is the first method that can do this for multiple types of data at once. Instead of just guessing how big the suitcase should be, it learns to separate the "shared" items (the things both cameras saw) from the "private" items (the things only one camera saw). In their tests, FiGuRO successfully figured out the true complexity of simulated data, even when the data was messy, noisy, or had very different amounts of information in each part. It didn't need complicated extra rules to force the data to separate; the separation happened naturally because the tool was so good at finding the most efficient way to pack the suitcase.

When the researchers tested FiGuRO on real-world data—like matching audio recordings of spoken numbers with images of those numbers, or combining radar and optical satellite photos—it worked just as well. It managed to strip away the noise and find the core "shared" signal, which helped computers recognize patterns better than before. The paper suggests that this approach is robust and reliable, working well across different types of data without needing to be tweaked constantly. It's like giving AI a pair of glasses that helps it see exactly how complex a situation really is, separating the signal from the noise so it can learn faster and smarter.