Unverified50% confidenceFactExact time
2018年Wachter等人将反事实解释形式化为可计算的损失函数：最小化特征变化量的同时约束模型输出翻转到期望类别
1
Sources
50%
Confidence
Long-term
Relevance
8/24/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedModern quantization schemes such as GPTQ and AWQ introduce layer-by-layer calibration and error compensation mechanisms to keep perplexity loss within 1–2%71% similarUnverifiedThe algorithm's loss function consists of three components: Focal Loss (positive-negative sample balancing), similarity loss (adjacent point consistency constraint), and shape loss (second-order smoothness constraint).70% similarUnverifiedSanturkar等人2018年的研究指出BN的真正作用可能在于平滑损失曲面而非减少内部协变量偏移69% similarUnverified香农的率失真理论证明，在给定比特率约束下，任何有损压缩都存在不可逾越的失真下界69% similarUnverifiedThe choice of bottleneck dimensionality is essentially a trade-off between information compression rate and reconstruction quality, corresponding to rate-distortion theory; Shannon proved in 1959 that there exists a theoretical minimum coding rate lower bound for a given distortion tolerance.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/793382API
curl https://kongchang.com/api/v1/knowledge/claims/793382MCP
get_claim(id=793382)