Machine learning models now have to work on messy, real-world data instead of textbook examples

Researchers found that existing machine learning systems break when new categories of data arrive at irregular intervals instead of in neat batches — which is how the real world actually works. The paper proposes fixes to make these systems more robust when data shows up unpredictably, rather than in the controlled conditions researchers have been testing.

Machine learning labs have spent years building systems that work perfectly in controlled experiments where new classes arrive in equal-sized, predictable batches. None of that happens in reality. A real system might see five new categories one week, then fifty the next week, then none for a month. This paper documents that existing methods fail badly under those conditions, and proposes concrete fixes. The gap between lab conditions and actual deployment has been obvious for years; this work makes the failure measurable and offers a path to patch it.