Good question Chris.

The reason why we don’t use the same neural network that we used for encoding to decode is

• To predict the output and compare it with the input data to calculate the loss. The output layer needs to predict the probability of an output which needs to either 0 or 1 and hence we use sigmoid activation function.
• Hidden layers for the encoder and decoder uses relu activation function as we want non linearity.

A sample code for AE would be something like this

input_img = Input(shape=(784,))
encoded = Dense(128, activation=’relu’)(input_img)
encoded = Dense(64, activation=’relu’)(encoded)
encoded = Dense(32, activation=’relu’)(encoded)

decoded = Dense(64, activation=’relu’)(encoded)
decoded = Dense(128, activation=’relu’)(decoded)
decoded = Dense(784, activation=’sigmoid’)(decoded)

Autoencoders are used to extract the latent feature representation so we also store the latent feature into a pickle file

I also plan to post AE code this week

For more on the AE refer to

Please let me know if this explanation was helpful