diff --git a/chapter_computational_graph/components_of_computational_graph.md b/chapter_computational_graph/components_of_computational_graph.md
index 4090e50..65073de 100644
--- a/chapter_computational_graph/components_of_computational_graph.md
+++ b/chapter_computational_graph/components_of_computational_graph.md
@@ -153,7 +153,7 @@ $$
 \nabla\boldsymbol{W_2} &= \boldsymbol{X_2}^\top\nabla\boldsymbol{Y}   \\
 \nabla\boldsymbol{X_1} &= \nabla\boldsymbol{X_2}\boldsymbol{W_1}^\top = (\nabla\boldsymbol{Y}\boldsymbol{W_2}^\top)\boldsymbol{W_1}^\top   \\
 \nabla\boldsymbol{W_1} &= \boldsymbol{X_1}^\top\nabla\boldsymbol{X_2} = \boldsymbol{X_1}^\top(\nabla\boldsymbol{Y}\boldsymbol{W_2}^\top)  \\
-\nabla\boldsymbol{Y} &= \nabla\boldsymbol{X_1}\boldsymbol{W}^\top = ((\nabla\boldsymbol{Y}\boldsymbol{W_2}^\top)\boldsymbol{W_1}^\top)\boldsymbol{W}^\top   \\
+\nabla\boldsymbol{X} &= \nabla\boldsymbol{X_1}\boldsymbol{W}^\top = ((\nabla\boldsymbol{Y}\boldsymbol{W_2}^\top)\boldsymbol{W_1}^\top)\boldsymbol{W}^\top   \\
 \nabla\boldsymbol{W} &= \boldsymbol{X}^\top\nabla\boldsymbol{X_1} = \boldsymbol{X}^\top((\nabla\boldsymbol{Y}\boldsymbol{W_2}^\top)\boldsymbol{W_1}^\top)
 \end{aligned}
 $$
@@ -162,12 +162,12 @@ $$
 
 根据上述公式我们可以得出循环控制的反向梯度计算过程如下，在下面代码中伪变量的前缀*grad*代表变量梯度变量，*transpose*代表矩阵转置算子。
 ```python
-grad_Y2 = matmul(grad_Y3, transpose(W2))
-grad_W2 = matmul(transpose(Y2), grad_Y3)
-grad_Y1 = matmul(grad_Y2, transpose(W1))
-grad_W1 = matmul(transpose(Y1), grad_Y2)
-grad_Y = matmul(grad_Y1, transpose(W))
-grad_W = matmul(transpose(Y), grad_Y1)
+grad_X2 = matmul(grad_Y, transpose(W2))
+grad_W2 = matmul(transpose(X2), grad_Y)
+grad_X1 = matmul(grad_X2, transpose(W1))
+grad_W1 = matmul(transpose(X1), grad_X2)
+grad_X = matmul(grad_X1, transpose(W))
+grad_W = matmul(transpose(X), grad_X1)
 ```
 结合公式、代码以及 :numref:`chain`我们可以看出，在反向传播过程中使用到前向传播的中间变量。因此保存网络中间层输出状态和中间变量，尽管占用了部分内存但能够复用计算结果，达到了提高反向传播计算效率的目的。