Do Vision CIL Methods Transfer to Protein? A Diagnostic Benchmark of Frozen Protein Encoders
Researchers tested nine class-incremental learning methods from vision tasks to see if they work for frozen protein representations. The study found that these methods do not transfer reliably across different protein encoders, regardless of how the data is split. Failures stem from the representation's unique geometry and changing statistical assumptions, suggesting future protein CIL needs encoder-aware designs.