Unverified50% confidenceFactExact time
The common premise of Xavier and Kaiming initialization is that the network structure is regular and the activation function is known and fixed; once the layer structure is custom or the activation function changes dynamically, the closed-form formulas need to be re-derived.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedXavier initialization (Glorot initialization) assumes that the activation function is approximately linear and symmetric about the origin, setting each layer's weight variance proportional to the inverse of the sum of input and output dimensions, so that the signal variance remains roughly constant during forward and backward propagation.66% similarUnverifiedThe paper's core theoretical contribution is a perturbation result: if the equilibrium selection is stable and the Gaussian transition model has uniform accuracy in the KL divergence sense, the induced value function will converge uniformly to the true value on the specified region, and the strategies approach approximately Nash behavior.66% similarUnverifiedThe existence of functional localisation provides theoretical support for directions such as parameter-efficient fine-tuning and modular editing.65% similarUnverifiedXavier初始化适用于Sigmoid或Tanh等饱和激活函数，根据输入输出维度调整方差范围64% similarUnverifiedArchitectures with recurrence can theoretically express iterative algorithms by repeatedly applying the same transformation, whereas fixed-depth feedforward networks have difficulty naturally expressing this type of computation.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1027618API
curl https://kongchang.com/api/v1/knowledge/claims/1027618MCP
get_claim(id=1027618)