知识卡片
小批量梯度用批大小归一化
若目标为N个样本损失的平均,均匀抽取B个样本时,通常以ĝ=(1/B)Σ批内∇L估计完整平均梯度。原书10.9和算法10.6印成1/N,会将期望缩小B/N,不能称为同尺度的平均估计;实现时须明确求和或求平均的约定。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
原书出处 · 1 处
- · 查看原书依据
For large N, computing this sum is very expensive. Suppose instead we randomly subsample (without replacement) a batch of terms from this sum, , where B is the batch size. We then compute an estimate of the total gradient as the average gradient over this batch as follows:
参考来源
- 位置:《Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)》第10章《基于梯度的学习算法》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Foundations of Computer Vision (Antonio Torralba, Phillip Isola etc.)/_epub-src/OPS/xhtml/c010.xhtml)
- 结论依据:摘录中关于小批量梯度用批大小归一化的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:For large N, computing this sum is very expensive. Suppose instead we randomly subsample (without replacement) a batch of terms from this sum, , where B is the batch size. We then compute an estimate of the total gradient as the average gradient over this batch as follows: