Unverified50% confidenceFactExact time
Memory bank methods (such as PatchCore and PaDiM) store a set of normal sample features during training and determine anomalies at inference via nearest-neighbor retrieval, with storage cost growing linearly with the training set size.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSWE-bench上已有多个研究团队记录到「patch记忆」现象，即模型通过识别问题描述特征检索训练时见过的补丁75% similarVerifiedFeature Stores address the issue of feature computation consistency between training time and inference time, known as training-serving skew.75% similarUnverifiedThe specialised predictor stores the entire finite topic mixture distribution, which amounts to memorizing the topic structure that appears in training.74% similarUnverifiedUpdating parametric memory requires fine-tuning or continual pre-training, both of which require additional training processes and computational resources.74% similarUnverifiedWhen saving checkpoints during distributed training, the data loader state and random seeds must also be saved; otherwise, after recovery the data order and augmentation results will be inconsistent, making training irreproducible.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1129020API
curl https://kongchang.com/api/v1/knowledge/claims/1129020MCP
get_claim(id=1129020)