Unverified50% confidenceFactExact time
反向传播一次前向传播加一次反向传播即可得到所有参数的梯度，计算量约为前向传播的2-3倍，与参数数量无关
1
Sources
50%
Confidence
Long-term
Relevance
8/13/2026
First Seen
Sources
Related Claims
UnverifiedTraditional gradient checkpointing trades memory savings by discarding some intermediate results from the forward pass and recomputing them during backpropagation, at the cost of approximately 30% additional computational overhead.61% similarUnverified在反向模式下，当模型参数数量高达数百万时，计算梯度的成本仅与一次前向计算相当61% similarUnverified自动微分在保持机器精度的同时，计算开销仅为前向传播的常数倍（通常2-5倍）57% similarUnverifiedIn High mode, the token consumption per chain-of-thought reasoning call may be 10 to 50 times that of Instant mode.56% similarUnverifiedLLM每次调用可建模为伯努利试验，N次调用通过次数服从二项分布B(N, p)，通过率95%置信区间约为p̂ ± 1.96×√(p̂(1-p̂)/N)56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/742244API
curl https://kongchang.com/api/v1/knowledge/claims/742244MCP
get_claim(id=742244)