Unverified50% confidenceCautionExact time
Never scale a node to exactly zero; a degenerate transform matrix disrupts culling and transform calculations, so scaling should be clamped to at least 0.02 and nearly flattened nodes hidden.
1
Sources
50%
Confidence
Long-term
Relevance
10/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFlattening a matrix and feeding it to an MLP results in a non-fixed input dimension due to variable row counts, and breaks permutation invariance.75% similarUnverifiedThe residual structure makes F=0 the default option—as long as the F(x) weights are close to zero, the residual block degenerates into an identity mapping, thereby resolving the optimization difficulty of SGD in learning identity mappings.72% similarUnverifiedThe eigenvalues of orthogonal matrix initialization have a modulus of 1, which neither grows nor decays exponentially under repeated matrix multiplication, effectively mitigating vanishing and exploding gradients.70% similarUnverified为Dice损失的分子分母添加平滑常数（通常取1.0）可以有效避免分母接近零时的数值不稳定70% similarUnverifiedIn the Decoder scenario, a Mask matrix must be introduced, assigning negative infinity to positions after the current word, which approach zero after Softmax, producing a causal generation effect.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1054334API
curl https://kongchang.com/api/v1/knowledge/claims/1054334MCP
get_claim(id=1054334)