Unverified50% confidenceFactExact time
Because neural networks have redundancy and compensatory mechanisms, ablating a small number of neurons often has limited effect, and multi-neuron or pathway-level analysis is needed.
1
Sources
50%
Confidence
Long-term
Relevance
10/28/2026
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79% similarUnverified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76% 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74% similarVerifiedAfter neuron alignment, the loss landscape between two networks often exhibits a linear pattern, allowing interpolated models to maintain considerable performance, whereas naive averaging typically falls into high-loss barrier regions.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% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1109180API
curl https://kongchang.com/api/v1/knowledge/claims/1109180MCP
get_claim(id=1109180)