Skip to content

Commit

Permalink
fix sentence
Browse files Browse the repository at this point in the history
Former-commit-id: 70e924cfe3bf6b3790dbfb31ec299aaab35039ec
  • Loading branch information
liber145 committed Jan 5, 2017
1 parent 0d82abc commit 2befafd
Show file tree
Hide file tree
Showing 6 changed files with 230 additions and 225 deletions.
239 changes: 121 additions & 118 deletions Chapter15/representation_learning.tex

Large diffs are not rendered by default.

188 changes: 94 additions & 94 deletions Chapter18/confronting_the_partition_function.tex

Large diffs are not rendered by default.

6 changes: 3 additions & 3 deletions Chapter2/linear_algebra.tex
Original file line number Diff line number Diff line change
Expand Up @@ -78,7 +78,7 @@ \section{标量,向量,矩阵和张量}
\item \firstgls{tensor}:在某些情况下,我们会讨论不只两维坐标的数组。
一般地,一组数组中的元素分布在若干维坐标的规则网格中,我们将其称之为张量。
我们使用这种字体$\TSA$来表示张量``A''
张量$\TSA$中坐标为$(i,j,k)$的元素记作$\textsf{\emph{A}}_{i,j,k}$
张量$\TSA$中坐标为$(i,j,k)$的元素记作$\TEA_{i,j,k}$
\end{itemize}


Expand Down Expand Up @@ -531,7 +531,7 @@ \section{\glsentrytext{eigendecomposition}}
从这个表示中我们可以获得一些有用的信息,比如$12$不能被$5$整除,或者$12$的倍数可以被$3$整除。


正如我们可以通过分解质因数来发现一些关于整数的真实性质,我们也可以通过分解矩阵来获取矩阵表示成数组元素时不明显的函数性质
正如我们可以通过分解质因数来发现整数的一些内在性质,我们也可以通过分解矩阵来发现矩阵表示成数组元素时不明显的函数性质


\firstgls{eigendecomposition}是使用最广的矩阵分解之一,即我们将矩阵分解成一组特征向量和特征值。
Expand All @@ -555,7 +555,7 @@ \section{\glsentrytext{eigendecomposition}}
类似地,我们也可以将特征值连接成一个向量$\Vlambda = [\lambda_1, \dots , \lambda_n]^\top$
因此$\MA$\firstgls{eigendecomposition}可以记作
\begin{equation}
\MA = \MV \text{diag}(\mathbf{\Vlambda}) \MV^{-1}.
\MA = \MV \text{diag}(\Vlambda) \MV^{-1}.
\end{equation}


Expand Down
8 changes: 3 additions & 5 deletions Chapter8/optimization_for_training_deep_models.tex
Original file line number Diff line number Diff line change
Expand Up @@ -98,11 +98,9 @@ \subsection{替代损失函数和\glsentrytext{early_stopping}}
这是因为即使$0-1$损失期望是零时,我们还能拉开不同类别的距离以改进分类器的鲁棒性,获得一个更强壮的,值得信赖的分类器。
因而,相较于简单地最小化训练集上的平均$0-1$损失,从训练数据中抽取了更多信息。

一般的优化和我们用于训练算法的优化的一个重要不同之处在于,训练算法通常不会停在一个局部最小处。
反之,机器学习通常会优化一个替代损失函数,并停止在基于\gls{early_stopping}(\secref{sec:early_stopping})的收敛条件满足的情况。
通常,\gls{early_stopping}是基于真实损失函数,如验证集上的$0-1$损失,并被设计用于将算法终止于过拟合发生之前。
训练过程通常终止于替代损失函数仍有较大的导数时,这和纯优化非常不同。
纯优化通常会考虑算法收敛到梯度很小的时候。
一般的优化和我们用于训练算法的优化的一个重要不同:训练算法通常不收敛在局部极小。
反之,机器学习通常优化替代损失函数,并可能基于\secref{sec:early_stopping}的收敛条件\gls{early_stopping}。
通常,\gls{early_stopping}使用真实损失函数,如验证集上的$0-1$损失函数,并在过拟合发生之前终止;与纯优化不同的是,\gls{early_stopping}时替代损失函数仍然有较大的导数,而纯优化终止时导数较小。

\subsection{批算法和\gls{minibatch}算法}
\label{sec:batch_and_minibatch_algorithms}
Expand Down
4 changes: 4 additions & 0 deletions acknowledgments_github.md
Original file line number Diff line number Diff line change
Expand Up @@ -144,3 +144,7 @@
- @endymecy ==> Chapter14, 语句问题


2017年1月5日
---------------
- @tonyzeng2016 ==> Chapter8, 8.1.2最后一段翻译建议
- @zdx3578 ==> Chapter18, "intractable"翻译建议
10 changes: 5 additions & 5 deletions terminology.tex
Original file line number Diff line number Diff line change
Expand Up @@ -645,7 +645,7 @@

\newglossaryentry{estimator}
{
name=估计量,
name=估计,
description={estimator},
sort={estimator},
}
Expand Down Expand Up @@ -2177,7 +2177,7 @@

\newglossaryentry{score_matching}
{
name=分数匹配,
name=得分匹配,
description={score matching},
sort={score matching},
}
Expand Down Expand Up @@ -4274,7 +4274,7 @@

\newglossaryentry{score}
{
name=分数,
name=得分,
description={score},
sort={score},
}
Expand Down Expand Up @@ -4632,7 +4632,7 @@

\newglossaryentry{spurious_modes}
{
name=伪造模式,
name=虚假模态,
description={spurious modes},
sort={spurious modes},
}
Expand Down Expand Up @@ -4677,7 +4677,7 @@

\newglossaryentry{GSM}
{
name=广义分数匹配,
name=广义得分匹配,
description={generalized score matching},
sort={generalized score matching},
symbol={GSM}
Expand Down

0 comments on commit 2befafd

Please sign in to comment.