Unverified50% confidenceFactExact time
The spectral bias of gradient descent — the eigenvalues of the loss function's Hessian matrix are larger in low-frequency directions, making low-frequency directions converge faster.
1
Sources
50%
Confidence
Long-term
Relevance
11/2/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedStiffness refers to the simultaneous presence of components with vastly different scales of change in a system, corresponding to situations where the Hessian matrix condition number of the loss surface is very large and the curvature varies greatly across different parameter directions.72% similarUnverifiedThe dynamics matrix H exhibits a graded structure: the eigenvalues corresponding to different frequency components decay in a predictable manner, with higher frequencies having smaller eigenvalues and correspondingly slower dynamical time scales.68% similarUnverifiedSparse rewards combined with long time spans cause gradient-based optimization signals to decay significantly during backpropagation.68% similarUnverifiedLow-rank decomposition and quantization compression are orthogonally complementary dimensions; quantizing the smaller matrices A and B separately results in smaller errors, and the combined compression ratio can reach several times that of either method alone66% similarUnverified低比特量化（Q4）会导致权重矩阵的秩降低，影响模型对低频但关键token（如JSON花括号、冒号）的预测置信度66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1125324API
curl https://kongchang.com/api/v1/knowledge/claims/1125324MCP
get_claim(id=1125324)