Unverified50% confidenceFactExact time
Self-supervised anomaly detection trains a reconstruction model using only normal samples, and at inference time treats regions with high reconstruction error as anomalies, naturally suited for scenarios with severe positive-negative sample imbalance.
1
Sources
50%
Confidence
Long-term
Relevance
10/9/2026
First Seen
Sources
Related Entities
Related Claims
Verified仅用成功轨迹训练的模型往往存在分布外泛化的严重缺陷，遇到工具返回错误时便丧失纠偏能力78% similarVerifiedAfter geometric transformation, the transformation parameters must be recorded and the predicted boxes must be inverse-transformed to align them; ignoring this step will cause semantic misalignment in the consistency loss and harm training effectiveness77% similarUnverifiedTampered model weights can have backdoors planted in them that produce preset erroneous outputs under specific trigger conditions, and are extremely difficult to detect through conventional testing.76% similarVerified若用含测试集的全量数据计算均值和标准差，会导致预处理泄漏，使评估指标虚高，无法真实反映模型泛化能力75% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/990938API
curl https://kongchang.com/api/v1/knowledge/claims/990938MCP
get_claim(id=990938)