diff --git a/README.md b/README.md
index 2c28bc58f..ea7b8577d 100644
--- a/README.md
+++ b/README.md
@@ -20,7 +20,7 @@ mlp = @mx.chain mx.Variable(:data)             =>
   mx.FullyConnected(name=:fc2, num_hidden=64)  =>
   mx.Activation(name=:relu2, act_type=:relu)   =>
   mx.FullyConnected(name=:fc3, num_hidden=10)  =>
-  mx.Softmax(name=:softmax)
+  mx.SoftmaxOutput(name=:softmax)
 
 # data provider
 batch_size = 100
diff --git a/docs/api/ndarray.rst b/docs/api/ndarray.rst
index 8ac5e9bda..05a3dccba 100644
--- a/docs/api/ndarray.rst
+++ b/docs/api/ndarray.rst
@@ -364,9 +364,9 @@ object (:class:`NDArray`) is returned. Otherwise, a tuple containing all the out
 
 Public APIs
 ^^^^^^^^^^^
-.. function:: choose_element(...)
+.. function:: choose_element_0index(...)
 
-   Choose one element from each line(row for python, column for R/Julia) in lhs according to index indicated by rhs
+   Choose one element from each line(row for python, column for R/Julia) in lhs according to index indicated by rhs. This function assume rhs uses 0-based index.
    
    :param lhs: Left operand to the function.
    :type lhs: NDArray
@@ -413,9 +413,42 @@ Public APIs
 
 
 
+.. function:: exp(...)
+
+   Take exp of the src
+   
+   :param src: Source input to the function
+   :type src: NDArray
+   
+
+
+
+
+.. function:: log(...)
+
+   Take log of the src
+   
+   :param src: Source input to the function
+   :type src: NDArray
+   
+
+
+
+
+.. function:: norm(...)
+
+   Take L2 norm of the src.The result will be ndarray of shape (1,) on the same device.
+   
+   :param src: Source input to the function
+   :type src: NDArray
+   
+
+
+
+
 .. function:: sqrt(...)
 
-   Take square root of the src
+   Take sqrt of the src
    
    :param src: Source input to the function
    :type src: NDArray
diff --git a/docs/api/symbol.rst b/docs/api/symbol.rst
index 2c7df712c..e01ecb359 100644
--- a/docs/api/symbol.rst
+++ b/docs/api/symbol.rst
@@ -143,7 +143,7 @@ Public APIs
    :type num_filter: int (non-negative), required
    
    
-   :param num_group: number of groups partition
+   :param num_group: Number of groups partition. This option is not supported by CuDNN, you can use SliceChannel to num_group,apply convolution and concat instead to achieve the same need.
    :type num_group: int (non-negative), optional, default=1
    
    
@@ -162,6 +162,57 @@ Public APIs
 
 
 
+.. function:: Deconvolution(...)
+
+   Apply deconvolution to input then add a bias.
+   
+   :param data: Input data to the DeconvolutionOp.
+   :type data: Symbol
+   
+   
+   :param weight: Weight matrix.
+   :type weight: Symbol
+   
+   
+   :param bias: Bias parameter.
+   :type bias: Symbol
+   
+   
+   :param kernel: deconvolution kernel size: (y, x)
+   :type kernel: Shape(tuple), required
+   
+   
+   :param stride: deconvolution stride: (y, x)
+   :type stride: Shape(tuple), optional, default=(1, 1)
+   
+   
+   :param pad: pad for deconvolution: (y, x)
+   :type pad: Shape(tuple), optional, default=(0, 0)
+   
+   
+   :param num_filter: deconvolution filter(channel) number
+   :type num_filter: int (non-negative), required
+   
+   
+   :param num_group: number of groups partition
+   :type num_group: int (non-negative), optional, default=1
+   
+   
+   :param workspace: Tmp workspace for deconvolution (MB)
+   :type workspace: long (non-negative), optional, default=512
+   
+   
+   :param no_bias: Whether to disable bias parameter.
+   :type no_bias: boolean, optional, default=True
+   
+   :param Base.Symbol name: The name of the symbol. (e.g. `:my_symbol`), optional.
+   
+   :return: the constructed :class:`Symbol`.
+   
+
+
+
+
 .. function:: Dropout(...)
 
    Apply dropout to input
@@ -412,7 +463,7 @@ Public APIs
 
 .. function:: Softmax(...)
 
-   Perform a softmax transformation on input.
+   DEPRECATED: Perform a softmax transformation on input. Please use SoftmaxOutput
    
    :param data: Input data to softmax.
    :type data: Symbol
@@ -433,9 +484,62 @@ Public APIs
 
 
 
+.. function:: SoftmaxOutput(...)
+
+   Perform a softmax transformation on input, backprop with logloss.
+   
+   :param data: Input data to softmax.
+   :type data: Symbol
+   
+   
+   :param grad_scale: Scale the gradient by a float factor
+   :type grad_scale: float, optional, default=1
+   
+   
+   :param multi_output: If set to true, for a (n,k,x_1,..,x_n) dimensionalinput tensor, softmax will generate n*x_1*...*x_n output, eachhas k classes
+   :type multi_output: boolean, optional, default=False
+   
+   :param Base.Symbol name: The name of the symbol. (e.g. `:my_symbol`), optional.
+   
+   :return: the constructed :class:`Symbol`.
+   
+
+
+
+
+.. function:: exp(...)
+
+   Take exp of the src
+   
+   :param src: Source symbolic input to the function
+   :type src: Symbol
+   
+   :param Base.Symbol name: The name of the symbol. (e.g. `:my_symbol`), optional.
+   
+   :return: the constructed :class:`Symbol`.
+   
+
+
+
+
+.. function:: log(...)
+
+   Take log of the src
+   
+   :param src: Source symbolic input to the function
+   :type src: Symbol
+   
+   :param Base.Symbol name: The name of the symbol. (e.g. `:my_symbol`), optional.
+   
+   :return: the constructed :class:`Symbol`.
+   
+
+
+
+
 .. function:: sqrt(...)
 
-   Take square root of the src
+   Take sqrt of the src
    
    :param src: Source symbolic input to the function
    :type src: Symbol
@@ -505,6 +609,25 @@ Internal APIs
 
 
 
+.. function:: _Native(...)
+
+   Stub for implementing an operator implemented in native frontend language.
+   
+   :param info: 
+   :type info: , required
+   
+   
+   :param need_top_grad: Whether this layer needs out grad for backward. Should be false for loss layers.
+   :type need_top_grad: boolean, optional, default=True
+   
+   :param Base.Symbol name: The name of the symbol. (e.g. `:my_symbol`), optional.
+   
+   :return: the constructed :class:`Symbol`.
+   
+
+
+
+
 .. function:: _Plus(...)
 
    Perform an elementwise plus.
diff --git a/docs/tutorial/mnist.rst b/docs/tutorial/mnist.rst
index 5fe21cde1..fc2e548dd 100644
--- a/docs/tutorial/mnist.rst
+++ b/docs/tutorial/mnist.rst
@@ -41,11 +41,11 @@ Note each composition we take the previous symbol as the `data` argument, formin
    Input --> 128 units (ReLU) --> 64 units (ReLU) --> 10 units
 
 where the last 10 units correspond to the 10 output classes (digits 0,...,9). We
-then add a final ``Softmax`` operation to turn the 10-dimensional prediction to proper probability values for the 10 classes:
+then add a final :class:`SoftmaxOutput` operation to turn the 10-dimensional prediction to proper probability values for the 10 classes:
 
 .. code-block:: julia
 
-   mlp  = mx.Softmax(data = fc3, name=:softmax)
+   mlp  = mx.SoftmaxOutput(data = fc3, name=:softmax)
 
 As we can see, the MLP is just a chain of layers. For this case, we can also use
 the ``mx.chain`` macro. The same architecture above can be defined as
@@ -58,7 +58,7 @@ the ``mx.chain`` macro. The same architecture above can be defined as
      mx.FullyConnected(name=:fc2, num_hidden=64)  =>
      mx.Activation(name=:relu2, act_type=:relu)   =>
      mx.FullyConnected(name=:fc3, num_hidden=10)  =>
-     mx.Softmax(name=:softmax)
+     mx.SoftmaxOutput(name=:softmax)
 
 After defining the architecture, we are ready to load the MNIST data. MXNet.jl
 provide built-in data providers for the MNIST dataset, which could automatically
diff --git a/examples/mnist/mlp.jl b/examples/mnist/mlp.jl
index b0703c56e..05d008d52 100644
--- a/examples/mnist/mlp.jl
+++ b/examples/mnist/mlp.jl
@@ -11,7 +11,7 @@ using MXNet
 # fc2  = mx.FullyConnected(data = act1, name=:fc2, num_hidden=64)
 # act2 = mx.Activation(data = fc2, name=:relu2, act_type=:relu)
 # fc3  = mx.FullyConnected(data = act2, name=:fc3, num_hidden=10)
-# mlp  = mx.Softmax(data = fc3, name=:softmax)
+# mlp  = mx.SoftmaxOutput(data = fc3, name=:softmax)
 
 #-- Option 2: using the mx.chain macro
 mlp = @mx.chain mx.Variable(:data)             =>