Based on the size of the hidden layer, AEs can be classified into two types, undercomplete AEs and overcomplete AEs:

• Undercomplete AE: If the AE simply learns to copy the input to the output, then it is not useful. The idea is to produce a concise representation as the output of the encoder, and this concise representation should consist of the most useful features of the input. The amount of conciseness achieved by the input layer is governed by the number of neurons or nodes that we use in the latent space representation. This can be set as a parameter while building the AE. If the number of neurons is set to fewer dimensions than that of the input features, then the AE is forced to learn most of the key features of the input data. The architecture where the number of neurons in latent space is less than that of input dimensions is called an undercomplete AE.
• Overcomplete AE: It is possible to represent the number of neurons in latent space as equal to or more than that of the input dimensions. This kind of architecture is termed an overcomplete AE.  In this case, the AE does not learn anything and simply copies the input to the latent space, which in turn is propagated through to the decoder.

Apart from the number of neurons in the latent space, the following are some of the other parameters that can be used in an AE architecture:

• Number of layers in the encoder and decoder: The depth of the encoder and decoder can be set to any number. Generally, in a mirrored architecture of encoder and decoder, the number of layers is set as the same number. The last figure is an illustration showing the AE with two layers, excluding the input and output, in both the encoder and decoder.
• Number of neurons per layer in encoder and decoder: The number of neurons decreases with each layer in an encoder and it increases with each layer in a decoder. The neurons in layers of encoders and decoders are symmetric.
• Loss function: Loss functions such as MSE or cross-entropy are used by AEs to learn the weights during backpropagation. If the input is in the range of (0,1), then cross-entropy is used as metric, otherwise MSE is used.