Unverified50% confidenceFactExact time
神经网络存在对抗样本脆弱性，微小的人类不可察觉的输入扰动可能导致模型输出剧烈变化
1
Sources
50%
Confidence
Long-term
Relevance
8/27/2026
First Seen
Sources
Related Entities
Related Claims
Unverified叠加现象指神经网络倾向于用远少于特征数量的神经元同时编码大量语义特征，与传统一个神经元一个特征的直觉相悖78% similarUnverified深度神经网络天然倾向于过拟合训练分布，对分布外样本的处理能力往往大幅下降75% similarUnverifiedAn intermittently faulty memory module may alternately produce correct and incorrect computation results, causing sporadic deviations in gradient calculations during neural network training, affecting model convergence without triggering hardware alerts.74% similarUnverified仅无约束地最大化SWD会导致神经网络扭曲数据结构得到退化解，因此需引入几何约束在分离性与保真度之间权衡73% similarUnverifiedVanishing gradient refers to the exponential decay of gradient values during backpropagation in deep neural networks, causing parameters near the input layer to receive almost no updates.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/810310API
curl https://kongchang.com/api/v1/knowledge/claims/810310MCP
get_claim(id=810310)