<table width="600" border="1">
<thead>
<tr>
<th>English Terminology</th>
<th>中文术语</th>
</tr>
</thead>
<tbody>
<tr>
<td>neural networks</td>
<td>神经网络</td>
</tr>
<tr>
<td>activation function</td>
<td>激活函数</td>
</tr>
<tr>
<td>hyperbolic tangent</td>
<td>双曲正切函数</td>
</tr>
<tr>
<td>bias units</td>
<td>偏置项</td>
</tr>
<tr>
<td>activation</td>
<td>激活值</td>
</tr>
<tr>
<td>forward propagation</td>
<td>前向传播</td>
</tr>
<tr>
<td>feedforward neural network</td>
<td>前馈神经网络</td>
</tr>
<tr>
<td>Backpropagation Algorithm</td>
<td>反向传播算法</td>
</tr>
<tr>
<td>(batch) gradient descent</td>
<td>（批量）梯度下降法</td>
</tr>
<tr>
<td>(overall) cost function</td>
<td>（整体）代价函数</td>
</tr>
<tr>
<td>squared-error</td>
<td>方差</td>
</tr>
<tr>
<td>average sum-of-squares error</td>
<td>均方差</td>
</tr>
<tr>
<td>regularization term</td>
<td>规则化项</td>
</tr>
<tr>
<td>weight decay</td>
<td>权重衰减</td>
</tr>
<tr>
<td>bias terms</td>
<td>偏置项</td>
</tr>
<tr>
<td>Bayesian regularization method</td>
<td>贝叶斯规则化方法</td>
</tr>
<tr>
<td>Gaussian prior</td>
<td>高斯先验概率</td>
</tr>
<tr>
<td>MAP</td>
<td>极大后验估计</td>
</tr>
<tr>
<td>maximum likelihood estimation</td>
<td>极大似然估计</td>
</tr>
<tr>
<td>activation function</td>
<td>激活函数</td>
</tr>
<tr>
<td>tanh function</td>
<td>双曲正切函数</td>
</tr>
<tr>
<td>non-convex function</td>
<td>非凸函数</td>
</tr>
<tr>
<td>hidden (layer) units</td>
<td>隐藏层单元</td>
</tr>
<tr>
<td>symmetry breaking</td>
<td>对称失效</td>
</tr>
<tr>
<td>learning rate</td>
<td>学习速率</td>
</tr>
<tr>
<td>forward pass</td>
<td>前向传导</td>
</tr>
<tr>
<td>hypothesis</td>
<td>假设值</td>
</tr>
<tr>
<td>error term</td>
<td>残差</td>
</tr>
<tr>
<td>weighted average</td>
<td>加权平均值</td>
</tr>
<tr>
<td>feedforward pass</td>
<td>前馈传导</td>
</tr>
<tr>
<td>Hadamard product</td>
<td>阿达马乘积</td>
</tr>
<tr>
<td>forward propagation</td>
<td>前向传播</td>
</tr>
<tr>
<td>off-by-one error</td>
<td>缺位错误</td>
</tr>
<tr>
<td>bias term</td>
<td>偏置项</td>
</tr>
<tr>
<td>numerically checking</td>
<td>数值检验</td>
</tr>
<tr>
<td>numerical roundoff errors</td>
<td>数值舍入误差</td>
</tr>
<tr>
<td>significant digits</td>
<td>有效数字</td>
</tr>
<tr>
<td>unrolling</td>
<td>组合扩展</td>
</tr>
<tr>
<td>learning rate</td>
<td>学习速率</td>
</tr>
<tr>
<td>Hessian matrix Hessian</td>
<td>矩阵</td>
</tr>
<tr>
<td>Newton's method</td>
<td>牛顿法</td>
</tr>
<tr>
<td>conjugate gradient</td>
<td>共轭梯度</td>
</tr>
<tr>
<td>step-size</td>
<td>步长值</td>
</tr>
<tr>
<td>Autoencoders</td>
<td>自编码算法</td>
</tr>
<tr>
<td>Sparsity</td>
<td>稀疏性</td>
</tr>
<tr>
<td>neural networks</td>
<td>神经网络</td>
</tr>
<tr>
<td>supervised learning</td>
<td>监督学习</td>
</tr>
<tr>
<td>unsupervised learning</td>
<td>无监督学习</td>
</tr>
<tr>
<td>hidden units</td>
<td>隐藏神经元</td>
</tr>
<tr>
<td>the pixel intensity value</td>
<td>像素灰度值</td>
</tr>
<tr>
<td>IID</td>
<td>独立同分布</td>
</tr>
<tr>
<td>PCA</td>
<td>主元分析</td>
</tr>
<tr>
<td>active</td>
<td>激活</td>
</tr>
<tr>
<td>inactive</td>
<td>抑制</td>
</tr>
<tr>
<td>activation function</td>
<td>激活函数</td>
</tr>
<tr>
<td>activation</td>
<td>激活度</td>
</tr>
<tr>
<td>the average activation</td>
<td>平均活跃度</td>
</tr>
<tr>
<td>sparsity parameter</td>
<td>稀疏性参数</td>
</tr>
<tr>
<td>penalty term</td>
<td>惩罚因子</td>
</tr>
<tr>
<td>KL divergence</td>
<td>KL 散度</td>
</tr>
<tr>
<td>Bernoulli random variable</td>
<td>伯努利随机变量</td>
</tr>
<tr>
<td>overall cost function</td>
<td>总体代价函数</td>
</tr>
<tr>
<td>backpropagation</td>
<td>后向传播</td>
</tr>
<tr>
<td>forward pass</td>
<td>前向传播</td>
</tr>
<tr>
<td>gradient descent</td>
<td>梯度下降</td>
</tr>
<tr>
<td>the objective</td>
<td>目标函数</td>
</tr>
<tr>
<td>the derivative checking method</td>
<td>梯度验证方法</td>
</tr>
<tr>
<td>Visualizing</td>
<td>可视化</td>
</tr>
<tr>
<td>Autoencoder</td>
<td>自编码器</td>
</tr>
<tr>
<td>hidden unit</td>
<td>隐藏单元</td>
</tr>
<tr>
<td>non-linear feature</td>
<td>非线性特征</td>
</tr>
<tr>
<td>activate</td>
<td>激励</td>
</tr>
<tr>
<td>trivial answer</td>
<td>平凡解</td>
</tr>
<tr>
<td>norm constrained</td>
<td>范数约束</td>
</tr>
<tr>
<td>sparse autoencoder</td>
<td>稀疏自编码器</td>
</tr>
<tr>
<td>norm bounded</td>
<td>有界范数</td>
</tr>
<tr>
<td>input domains</td>
<td>输入域</td>
</tr>
<tr>
<td>vectorization</td>
<td>矢量化</td>
</tr>
<tr>
<td>Logistic Regression</td>
<td>逻辑回归</td>
</tr>
<tr>
<td>batch gradient ascent</td>
<td>批量梯度上升法</td>
</tr>
<tr>
<td>intercept term</td>
<td>截距</td>
</tr>
<tr>
<td>the log likelihood</td>
<td>对数似然函数</td>
</tr>
<tr>
<td>derivative</td>
<td>导函数</td>
</tr>
<tr>
<td>gradient</td>
<td>梯度</td>
</tr>
<tr>
<td>vectorization</td>
<td>向量化</td>
</tr>
<tr>
<td>forward propagation</td>
<td>正向传播</td>
</tr>
<tr>
<td>backpropagation</td>
<td>反向传播</td>
</tr>
<tr>
<td>training examples</td>
<td>训练样本</td>
</tr>
<tr>
<td>activation function</td>
<td>激活函数</td>
</tr>
<tr>
<td>sparse autoencoder</td>
<td>稀疏自编码网络</td>
</tr>
<tr>
<td>sparsity penalty</td>
<td>稀疏惩罚</td>
</tr>
<tr>
<td>average firing rate</td>
<td>平均激活率</td>
</tr>
<tr>
<td>Principal Components Analysis</td>
<td>主成份分析</td>
</tr>
<tr>
<td>whitening</td>
<td>白化</td>
</tr>
<tr>
<td>intensity</td>
<td>亮度</td>
</tr>
<tr>
<td>mean</td>
<td>平均值</td>
</tr>
<tr>
<td>variance</td>
<td>方差</td>
</tr>
<tr>
<td>covariance matrix</td>
<td>协方差矩阵</td>
</tr>
<tr>
<td>basis</td>
<td>基</td>
</tr>
<tr>
<td>magnitude</td>
<td>幅值</td>
</tr>
<tr>
<td>stationarity</td>
<td>平稳性</td>
</tr>
<tr>
<td>normalization</td>
<td>归一化</td>
</tr>
<tr>
<td>eigenvector</td>
<td>特征向量</td>
</tr>
<tr>
<td>redundant</td>
<td>冗余</td>
</tr>
<tr>
<td>variance</td>
<td>方差</td>
</tr>
<tr>
<td>smoothing</td>
<td>平滑</td>
</tr>
<tr>
<td>dimensionality reduction</td>
<td>降维</td>
</tr>
<tr>
<td>regularization</td>
<td>正则化</td>
</tr>
<tr>
<td>reflection matrix</td>
<td>反射矩阵</td>
</tr>
<tr>
<td>decorrelation</td>
<td>去相关</td>
</tr>
<tr>
<td>Principal Components Analysis (PCA)</td>
<td>主成分分析</td>
</tr>
<tr>
<td>zero-mean</td>
<td>均值为零</td>
</tr>
<tr>
<td>mean value</td>
<td>均值</td>
</tr>
<tr>
<td>eigenvalue</td>
<td>特征值</td>
</tr>
<tr>
<td>symmetric positive semi-definite matrix</td>
<td>对称半正定矩阵</td>
</tr>
<tr>
<td>numerically reliable</td>
<td>数值计算上稳定</td>
</tr>
<tr>
<td>sorted in decreasing order</td>
<td>降序排列</td>
</tr>
<tr>
<td>singular value</td>
<td>奇异值</td>
</tr>
<tr>
<td>singular vector</td>
<td>奇异向量</td>
</tr>
<tr>
<td>vectorized implementation</td>
<td>向量化实现</td>
</tr>
<tr>
<td>diagonal</td>
<td>对角线</td>
</tr>
<tr>
<td>Softmax Regression</td>
<td>Softmax回归</td>
</tr>
<tr>
<td>supervised learning</td>
<td>有监督学习</td>
</tr>
<tr>
<td>unsupervised learning</td>
<td>无监督学习</td>
</tr>
<tr>
<td>deep learning</td>
<td>深度学习</td>
</tr>
<tr>
<td>logistic regression</td>
<td>logistic回归</td>
</tr>
<tr>
<td>intercept term</td>
<td>截距项</td>
</tr>
<tr>
<td>binary classification</td>
<td>二元分类</td>
</tr>
<tr>
<td>class labels</td>
<td>类型标记</td>
</tr>
<tr>
<td>hypothesis</td>
<td>估值函数/估计值</td>
</tr>
<tr>
<td>cost function</td>
<td>代价函数</td>
</tr>
<tr>
<td>multi-class classification</td>
<td>多元分类</td>
</tr>
<tr>
<td>weight decay</td>
<td>权重衰减</td>
</tr>
<tr>
<td>self-taught learning</td>
<td>自我学习/自学习</td>
</tr>
<tr>
<td>unsupervised feature learning</td>
<td>无监督特征学习</td>
</tr>
<tr>
<td>autoencoder</td>
<td>自编码器</td>
</tr>
<tr>
<td>semi-supervised learning</td>
<td>半监督学习</td>
</tr>
<tr>
<td>deep networks</td>
<td>深层网络</td>
</tr>
<tr>
<td>fine-tune</td>
<td>微调</td>
</tr>
<tr>
<td>unsupervised feature learning</td>
<td>非监督特征学习</td>
</tr>
<tr>
<td>pre-training</td>
<td>预训练</td>
</tr>
<tr>
<td>Deep Networks</td>
<td>深度网络</td>
</tr>
<tr>
<td>deep neural networks</td>
<td>深度神经网络</td>
</tr>
<tr>
<td>non-linear transformation</td>
<td>非线性变换</td>
</tr>
<tr>
<td>represent compactly</td>
<td>简洁地表达</td>
</tr>
<tr>
<td>part-whole decompositions</td>
<td>“部分-整体”的分解</td>
</tr>
<tr>
<td>parts of objects</td>
<td>目标的部件</td>
</tr>
<tr>
<td>highly non-convex optimization problem</td>
<td>高度非凸的优化问题</td>
</tr>
<tr>
<td>conjugate gradient</td>
<td>共轭梯度</td>
</tr>
<tr>
<td>diffusion of gradients</td>
<td>梯度的弥散</td>
</tr>
<tr>
<td>Greedy layer-wise training</td>
<td>逐层贪婪训练方法</td>
</tr>
<tr>
<td>autoencoder</td>
<td>自动编码器</td>
</tr>
<tr>
<td>Greedy layer-wise training</td>
<td>逐层贪婪训练法</td>
</tr>
<tr>
<td>Stacked autoencoder</td>
<td>栈式自编码神经网络</td>
</tr>
<tr>
<td>Fine-tuning</td>
<td>微调</td>
</tr>
<tr>
<td>Raw inputs</td>
<td>原始输入</td>
</tr>
<tr>
<td>Hierarchical grouping</td>
<td>层次型分组</td>
</tr>
<tr>
<td>Part-whole decomposition</td>
<td>部分-整体分解</td>
</tr>
<tr>
<td>First-order features</td>
<td>一阶特征</td>
</tr>
<tr>
<td>Second-order features</td>
<td>二阶特征</td>
</tr>
<tr>
<td>Higher-order features</td>
<td>更高阶特征</td>
</tr>
<tr>
<td>Linear Decoders</td>
<td>线性解码器</td>
</tr>
<tr>
<td>Sparse Autoencoder</td>
<td>稀疏自编码</td>
</tr>
<tr>
<td>input layer</td>
<td>输入层</td>
</tr>
<tr>
<td>hidden layer</td>
<td>隐含层</td>
</tr>
<tr>
<td>output layer</td>
<td>输出层</td>
</tr>
<tr>
<td>neuron</td>
<td>神经元</td>
</tr>
<tr>
<td>robust</td>
<td>鲁棒</td>
</tr>
<tr>
<td>sigmoid activation function</td>
<td>S型激励函数</td>
</tr>
<tr>
<td>tanh function</td>
<td>tanh激励函数</td>
</tr>
<tr>
<td>linear activation function</td>
<td>线性激励函数</td>
</tr>
<tr>
<td>identity activation function</td>
<td>恒等激励函数</td>
</tr>
<tr>
<td>hidden unit</td>
<td>隐单元</td>
</tr>
<tr>
<td>weight</td>
<td>权重</td>
</tr>
<tr>
<td>error term</td>
<td>偏差项</td>
</tr>
<tr>
<td>Full Connected Networks</td>
<td>全连接网络</td>
</tr>
<tr>
<td>Sparse Autoencoder</td>
<td>稀疏编码</td>
</tr>
<tr>
<td>Feedforward</td>
<td>前向输送</td>
</tr>
<tr>
<td>Backpropagation</td>
<td>反向传播</td>
</tr>
<tr>
<td>Locally Connected Networks</td>
<td>部分联通网络</td>
</tr>
<tr>
<td>Contiguous Groups</td>
<td>连接区域</td>
</tr>
<tr>
<td>Visual Cortex</td>
<td>视觉皮层</td>
</tr>
<tr>
<td>Convolution</td>
<td>卷积</td>
</tr>
<tr>
<td>Stationary</td>
<td>固有特征</td>
</tr>
<tr>
<td>Pool</td>
<td>池化</td>
</tr>
<tr>
<td>features</td>
<td>特征</td>
</tr>
<tr>
<td>example</td>
<td>样例</td>
</tr>
<tr>
<td>over-fitting</td>
<td>过拟合</td>
</tr>
<tr>
<td>translation invariant</td>
<td>平移不变性</td>
</tr>
<tr>
<td>pooling</td>
<td>池化</td>
</tr>
<tr>
<td>extract</td>
<td>提取</td>
</tr>
<tr>
<td>object detection</td>
<td>物体检测</td>
</tr>
<tr>
<td>DC component</td>
<td>直流分量</td>
</tr>
<tr>
<td>local mean subtraction</td>
<td>局部均值消减</td>
</tr>
<tr>
<td>sparse autoencoder</td>
<td>消减归一化</td>
</tr>
<tr>
<td>rescaling</td>
<td>缩放</td>
</tr>
<tr>
<td>per-example mean subtraction</td>
<td>逐样本均值消减</td>
</tr>
<tr>
<td>feature standardization</td>
<td>特征标准化</td>
</tr>
<tr>
<td>stationary</td>
<td>平稳</td>
</tr>
<tr>
<td>zero-mean</td>
<td>零均值化</td>
</tr>
<tr>
<td>low-pass filtering</td>
<td>低通滤波</td>
</tr>
<tr>
<td>reconstruction based models</td>
<td>基于重构的模型</td>
</tr>
<tr>
<td>RBMs</td>
<td>受限Boltzman机</td>
</tr>
<tr>
<td>k-Means</td>
<td>k-均值</td>
</tr>
<tr>
<td>long tail</td>
<td>长尾</td>
</tr>
<tr>
<td>loss function</td>
<td>损失函数</td>
</tr>
<tr>
<td>orthogonalization</td>
<td>正交化</td>
</tr>
<tr>
<td>Sparse Coding</td>
<td>稀疏编码</td>
</tr>
<tr>
<td>unsupervised method</td>
<td>无监督学习</td>
</tr>
<tr>
<td>over-complete bases</td>
<td>超完备基</td>
</tr>
<tr>
<td>degeneracy</td>
<td>退化</td>
</tr>
<tr>
<td>reconstruction term</td>
<td>重构项</td>
</tr>
<tr>
<td>sparsity penalty</td>
<td>稀疏惩罚项</td>
</tr>
<tr>
<td>norm</td>
<td>范式</td>
</tr>
<tr>
<td>generative model</td>
<td>生成模型</td>
</tr>
<tr>
<td>linear superposition</td>
<td>线性叠加</td>
</tr>
<tr>
<td>additive noise</td>
<td>加性噪声</td>
</tr>
<tr>
<td>basis feature vectors</td>
<td>特征基向量</td>
</tr>
<tr>
<td>the empirical distribution</td>
<td>经验分布函数</td>
</tr>
<tr>
<td>the log-likelihood</td>
<td>对数似然函数</td>
</tr>
<tr>
<td>Gaussian white noise</td>
<td>高斯白噪音</td>
</tr>
<tr>
<td>the prior distribution</td>
<td>先验分布</td>
</tr>
<tr>
<td>prior probability</td>
<td>先验概率</td>
</tr>
<tr>
<td>source features</td>
<td>源特征</td>
</tr>
<tr>
<td>the energy function</td>
<td>能量函数</td>
</tr>
<tr>
<td>regularized</td>
<td>正则化</td>
</tr>
<tr>
<td>least squares</td>
<td>最小二乘法</td>
</tr>
<tr>
<td>convex optimization software</td>
<td>凸优化软件</td>
</tr>
<tr>
<td>conjugate gradient methods</td>
<td>共轭梯度法</td>
</tr>
<tr>
<td>quadratic constraints</td>
<td>二次约束</td>
</tr>
<tr>
<td>the Lagrange dual</td>
<td>拉格朗日对偶函数</td>
</tr>
<tr>
<td>feedforward architectures</td>
<td>前馈结构算法</td>
</tr>
<tr>
<td>Independent Component Analysis</td>
<td>独立成分分析</td>
</tr>
<tr>
<td>Over-complete basis</td>
<td>超完备基</td>
</tr>
<tr>
<td>Orthonormal basis</td>
<td>标准正交基</td>
</tr>
<tr>
<td>Sparsity penalty</td>
<td>稀疏惩罚项</td>
</tr>
<tr>
<td>Under-complete basis</td>
<td>不完备基</td>
</tr>
<tr>
<td>Line-search algorithm</td>
<td>线搜索算法</td>
</tr>
<tr>
<td>Topographic cost term</td>
<td>拓扑代价项</td>
</tr>
</tbody>
</table>
来源：机器人网
