Galaxy Morphology Classification: Uncertainty Modeling and Out of Distribution Detection

The universe is filled with billions of galaxies, each a vast island of stars, gas, and dust. For nearly a century, astronomers have sorted these cosmic islands into families based on their shapes: some are smooth and round like giant balls of light, while others are flat spirals with winding arms. This sorting, known as morphological classification, is not just a matter of cataloging; the shape of a galaxy tells a story about its history, how it formed, and how it has changed over time. However, as modern telescopes capture images of the sky with unprecedented clarity and volume, the sheer number of galaxies has become too great for human eyes to examine one by one. To keep pace, scientists have turned to artificial intelligence, teaching computers to recognize these shapes automatically. But there is a catch: standard computer programs are often overconfident. When they encounter a galaxy that looks strange or unlike anything they have seen before, they will still force it into a known category and claim to be certain, potentially hiding rare or new discoveries behind a wall of false confidence.

A team of researchers at the Indian Institute of Technology Hyderabad set out to fix this problem by building a smarter way for computers to classify galaxies. They focused on a specific challenge called "out-of-distribution" detection, which is simply the ability to recognize when an image does not belong to the known groups. Using a massive collection of galaxy images from the Dark Energy Camera Legacy Survey, they trained a deep learning system to not only identify the nine standard shapes of galaxies but also to admit when it was unsure. The researchers tested three different approaches. The first was a standard method used as a baseline. The second introduced a new technique that measures how far a galaxy's features are from the center of its known group, rather than just guessing a category. The third combined this distance measurement with a method that runs the same image through the computer's brain multiple times with slight variations, allowing the system to gauge its own uncertainty.

The results showed that the new methods were far superior to the standard approach. The system using the distance-based technique became significantly better at spotting unusual galaxies, improving its ability to correctly identify them as "unknown" by nearly 90 percent compared to the baseline, while still correctly identifying the known shapes with 97 percent accuracy. Crucially, this system learned to be humble. When it encountered a galaxy that was difficult to classify or looked like a mix of different types, it did not force a label. Instead, it signaled uncertainty, often by showing that the galaxy was far away from the center of any known shape group. This behavior is vital for future sky surveys, where the goal is not just to count what we know, but to find the rare, the weird, and the previously unseen.

The researchers also found that combining the distance method with the multiple-pass technique made the system even more reliable. By running the image through the network many times, the computer could see if its answer changed slightly with each pass. If the answer wavered, the system knew it was dealing with something ambiguous. This approach reduced the error in the system's confidence scores by about 73 percent, meaning that when the computer said it was sure, it was genuinely sure, and when it said it was unsure, it was correctly identifying a difficult case. The study demonstrated that by changing how the computer measures similarity—focusing on geometric distance rather than just probability—it could create a clearer separation between familiar galaxies and those that are truly new or strange.

This work does not just improve the speed of classification; it changes the reliability of the entire process. In the past, an automated system might have confidently mislabeled a rare, merging galaxy as a common spiral, effectively erasing a unique cosmic event from the record. The new framework ensures that such anomalies are flagged for human review. The researchers verified that the computer was paying attention to the right parts of the images, such as spiral arms or central bulges, rather than random noise or background artifacts. When the system made a mistake, it was often because the galaxy itself was confusing, with features that blended together, and the system correctly reflected this confusion through higher uncertainty.

The implications for astronomy are significant. Upcoming telescopes will soon capture images of billions of galaxies, a volume that makes manual checking impossible. Having a tool that can automatically sort the knowns while reliably pointing out the unknowns is essential for discovering the next generation of cosmic phenomena. The researchers noted that while their method is highly effective, it does require more computing power to run the multiple checks needed for uncertainty. However, they argue that this cost is a small price to pay for the ability to trust the automated catalog and to ensure that no strange or rare galaxy slips through the cracks unnoticed. By teaching machines to recognize the limits of their own knowledge, astronomers can finally let the computers handle the routine work while focusing their own attention on the most mysterious and fascinating objects in the universe.