Create an ElementwiseLayer that computes Tanh of each input element:

Create an ElementwiseLayer that multiplies its input by a fixed constant:

Apply the layer to an input vector:

Create a ElementwiseLayer that computes a "hard sigmoid":

Apply the layer to a batch of inputs:

Plot the behavior of the layer:

Create an ElementwiseLayer that takes vectors of size 3:

Apply the layer to a single vector:

When applied, the layer will automatically thread over a batch of vectors:

Create an ElementwiseLayer that computes a Gaussian:

Apply the layer to a range of values and plot the result:

Train a 16-layer-deep, self-normalizing network described in "Self-Normalizing Neural Networks", G. Klambauer et. al., 2017, on the UCI Letter dataset. Obtain the data:

Self-normalizing nets assume that the input data has mean of 0 and variance of 1. Standardize the test and training data:

Verify that the sample mean and variance of the training data are 0 and 1, respectively:

Define a 16-layer, self-normalizing net with "AlphaDropout":

Train the net:

Obtain the accuracy:

Compare the accuracy against the "RandomForest" method in Classify:

Allow a classification network to deal with a "non-separable" problem. Create a synthetic training set consisting of points on a disk, separated into two classes by the circle r=0.5:

Create a layer composed of two LinearLayer layers, and a final transformation into a probability using an ElementwiseLayer:

Train the network on the data:

The net was not able to separate the two classes:

Because LinearLayer is an affine layer, stacking the two layers without an intervening nonlinearity is equivalent to using a single layer. A single line in the plane cannot separate the two classes, which is the level set of a single LinearLayer.

Train a similar net that has a Ramp nonlinearity between the two layers:

The net can now separate the classes:

Binary classification tasks require that the output of a net be a probability. ElementwiseLayer[LogisticSigmoid] can be used to take an arbitrary scalar and produce a value between 0 and 1. Create a net that takes a vector of length 2 and produces a binary prediction:

Train the net to decide if the first number in the vector is greater than the second:

Evaluate the net on several inputs:

The underlying output is a probability, which can be seen by disabling the "Boolean" decoder:

ElementwiseLayer is automatically used when an appropriate function is specified in a NetChain or NetGraph:

ElementwiseLayer cannot accept symbolic inputs:

Certain choices of f can produce failures for inputs outside their domain:

Certain functions are not supported directly:

Approximate the Zeta function using a rational function:

The approximation is good over the range (-10,10):

Construct an ElementwiseLayer using the approximation:

Measure the error of the approximation on specific inputs:

The network will fail when evaluated at a pole: