Unverified50% confidenceFactExact time
BatchNorm在推理时可以直接折叠进卷积的权重和偏置中，折叠后新权重 W' = γW/√(σ²+ε)，新偏置 b' = γ(b-μ)/√(σ²+ε) + β
1
Sources
50%
Confidence
Long-term
Relevance
7/24/2026
First Seen
Sources
AI Application Engineer Interview Guide: Core Topics in Model Quantization and Inference Deployment
redditr/learnmachinelearning7/9/2026
Related Claims
UnverifiedCollapse模型的状态更新公式为 h ← h − strength · (1 − cos(h, W)) · normalize(h − W)60% similarUnverified两个不可约表示的张量积l₁⊗l₂可分解为|l₁-l₂|到l₁+l₂之间所有整数阶不可约表示的直和，CG系数给出具体的基变换矩阵57% similarUnverifiedL2算法相对于基向量系数位长β达到O(n⁴·β·(n+β))的位复杂度，在β=O(n)时相当于O(n⁶)，而经典LLL约为O(n⁹)56% similarVerified混合专家架构MoE通过稀疏激活机制，推理时只激活模型总参数的1/8至1/455% similarUnverifiedAutomatic differentiation precisely applies the chain rule by decomposing a program into a sequence of elementary operations, avoiding both the rounding errors of numerical differentiation and the expression swell of symbolic differentiation.55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/611108API
curl https://kongchang.com/api/v1/knowledge/claims/611108MCP
get_claim(id=611108)