Unverified50% confidenceFactExact time
In neural network weight research, low-rank structure means that weight matrices often have a low effective rank, with their information concentrated in a few singular value directions.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
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蒸馏的核心洞察在于神经网络的软标签（类别概率分布）包含比one-hot硬标签更丰富的结构信息76% similarUnverified线性代数中的矩阵分解与特征值分解对应神经网络的权重空间分析74% similarUnverifiedBecause 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.72% similarUnverifiedNetworks that only memorize without generalizing show a much smaller degree of effective-rank compression.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1130935API
curl https://kongchang.com/api/v1/knowledge/claims/1130935MCP
get_claim(id=1130935)