Unverified50% confidenceFactExact time
MIRA's perturbation scale selection depends on mutual information estimation and may face estimation variance issues when batches are small or the number of classes is very large.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMutual information measures the degree of dependence between the prediction and the perturbed classifiers; if the prediction changes sharply under perturbation, the MI is small.80% similarUnverifiedThe higher the VC dimension, the more complex the region and the larger the estimation error may be; the larger the mutual information, the stronger the association between region selection and the sample, and the larger the estimation error.73% similarUnverifiedMIRA uses an entropy loss weighted by mutual-information confidence, reducing the weight of samples that are perturbation-sensitive or low-confidence.72% similarUnverifiedInstance normalization discards amplitude information, and amplitude may be a key discriminative feature in some classification tasks.72% similarUnverifiedThe impact of data quality on fine-tuning results is often far greater than that of hyperparameter tuning, and this is especially critical in scenarios with limited sample sizes.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1129530API
curl https://kongchang.com/api/v1/knowledge/claims/1129530MCP
get_claim(id=1129530)