Beginning Anomaly Detection Using

 What Are Autoencoders?

In the previous chapter, you learned about the basic functioning of a neural network. 

The basic concept is that a neural network essentially computes a weighted calculation 

of inputs to produce outputs. The inputs are in the input layer and the outputs are in 

the output layer and there are one or more hidden layers between the input and output 

layers. Back propagation is a technique used to train the network while trying to adjust 

the weights until the error is minimized. Autoencoders use this property of a neural 

network in a special way to accomplish some very efficient methods of training networks 

to learn normal behavior, thus helping to detect anomalies when they occur. Figure 

4-1

 

shows a typical neural network.

---

124

Autoencoders are neural networks that have the ability to discover low-dimensional 

representations of high-dimensional data and are able to reconstruct the input from the 

output. Autoencoders are made up of two pieces of the neural network, an encoder and 

a decoder. The encoder reduces the dimensionality of a high dimensional dataset to a 

low dimensional one whereas a decoder essentially expands the low-dimensional data 

to high-dimensional data. The goal of such a process is to try to reconstruct the original 

input. If the neural network is good, then there is a good chance of reconstructing the 

original input from the encoded data. This inherent principle is critical in building an 

anomaly detection module.

Note that autoencoders are not that great if you have training samples containing 

few dimensions/features at each input point. Autoencoders perform well for five or more 

dimensions. If you have just one dimension/feature then, as you can imagine, you are 

just doing a linear transformation, which is not useful.

Figure 4-1.  A typical neural network

Chapter 4   autoenCoders

---

125

Autoencoders are incredibly useful in many use cases. Some popular applications of 

autoencoders are

  1.  Training deep learning networks

 2. Compression

 3. Classification

  4.  Anomaly detection

  5.  Generative models