PEAKS: Selecting Key Training Examples Incrementally via Prediction Error Anchored by Kernel Similarity | Read Paper on Bytez