Unverified50% confidenceFactExact time
In the abstraction phase, H-GBP first constructs a coarse-grained graph approximation, transforming global errors that would otherwise need to propagate across a large number of nodes into problems solvable over short distances.
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGBP can rapidly smooth local errors between adjacent variables, but global errors spanning distant regions of the graph can only be corrected incrementally through long-distance message propagation, causing convergence speed to slow significantly on large-scale graphs.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65% similarUnverified在计算交叉熵损失时，softmax对正确标签位置i的梯度为p_i-1，对其他位置j的梯度为p_j，使lm_head输出层梯度密集64% similarUnverifiedHNSW通过构建多层图结构模拟跳表的层级跳跃特性，能在近似对数时间复杂度内完成近似最近邻查找，但内存占用相对较高且原生不支持量化压缩63% similarUnverifiedThe theoretical guarantees of Decoupled Descent rest on strong assumptions including stylized Gaussian mixture data, full-batch gradient descent, and high-dimensional asymptotic limits, and there is still a long way to go before it can be applied to large-scale models.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/983406API
curl https://kongchang.com/api/v1/knowledge/claims/983406MCP
get_claim(id=983406)