← 返回刚才阅读的位置

知识卡片

结构图卡 · 9261 · 生成模型:GPT-6

梯度计算与优化更新职责不同

理解等级:中级

反向传播按链式法则计算损失对参数的导数,优化器据梯度及其状态决定更新。导数是局部敏感度,不是可相加的错误归因;损失下降也不保证分类准确率每步上升。

结构图:

flowchart LR
 A[训练样本与当前参数] --> B[前向预测]
 B --> C[计算训练损失]
 C --> D[反向传播求梯度]
 D --> E[优化器决定更新]
 E --> F[下一步参数]
这张卡片看不懂?

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

      Backward pass: the objective of measuring the loss in Step 3 is to minimize it. Naturally, by minimizing the loss, the performance and accuracy of the model should increase. The next step in the training process is to do exactly that via the backward pass. In this stage, an optimization algorithm (usually gradient descent) is deployed which calculates the loss (from Step 3) with respect

    参考来源

    - 位置:《Machine Learning Fundamentals: Concepts, Models, and Applications》第09章《训练机器学习模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Machine Learning Fundamentals: Concepts, Models, and Applications/_epub-src/OEBPS/xhtml/23_chapter09.xhtml) - 结论依据:摘录中关于梯度计算与优化更新职责不同的定义、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:Backward pass: the objective of measuring the loss in Step 3 is to minimize it. Naturally, by minimizing the loss, the performance and accuracy of the model should increase. The next step in the training process is to do exactly that via the backward pass. In this stage, an optimization algorithm (usually gradient descent) is deployed which calculates the loss (from Step 3) with respect