Unverified50% confidenceFactExact time
The authors derived the eigenvalue decay rate of the associated integral operator for shallow networks (using RePU activation and zero-bias initialization).
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedAlthough over-parameterized networks have enough capacity to memorize random labels, the implicit regularization of gradient descent favors learning simple, consistent patterns first.67% similarUnverifiedTo improve generation accuracy, the research introduced two mechanisms: zero-displacement prediction and size-based grouping.67% similarUnverified反向传播机制要求整个计算图必须是可微分的，且在极深网络中存在梯度消失/爆炸问题66% similarUnverifiedThe authors derive a Gaussian posterior predictive formula, using a structured covariance representation to accelerate computation in order to reduce computational cost.66% similarUnverifiedThe author points out that the direct gradient path from the current token to historical information decays at a polynomial rate of 1/(t·ln t), far slower than the exponential decay of traditional RNNs.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1102475API
curl https://kongchang.com/api/v1/knowledge/claims/1102475MCP
get_claim(id=1102475)