Research collection
#noise-calibration
Research tagged noise-calibration: explanations, evidence, limitations and connections between papers.
2 results
A direct formula makes private-training noise calibration less conservative
A direct noise-calibration method bounds membership tests in random-allocation DP-SGD and computes an exact reference.
Why language models can overstate their progress
A statistical model and loop tests show why reused examples can make language-model progress scores too optimistic.