Autoencoders in Deep Learning Explained

Quick answer: Learn what autoencoders are, how the encoder-decoder architecture works, their use in dimensionality reduction and anomaly detection, with a simple example.

What is an Autoencoder?

An autoencoder is a type of neural network trained to reconstruct its own input, forcing it to learn a compressed, meaningful representation of the data in the process. It is a form of unsupervised learning, since it needs no labels beyond the input data itself.

The Encoder-Decoder Architecture

An autoencoder has two parts. The encoder compresses the input into a smaller representation, called the latent space or bottleneck. The decoder then reconstructs the original input from that compressed representation.

Input -> Encoder -> Compressed representation (bottleneck) -> Decoder -> Reconstructed output

A Simple Autoencoder in Keras

from tensorflow import keras
from tensorflow.keras import layers

input_dim = 784   # e.g. a flattened 28x28 image
encoding_dim = 32  # compressed representation size

input_layer = keras.Input(shape=(input_dim,))
encoded = layers.Dense(encoding_dim, activation='relu')(input_layer)
decoded = layers.Dense(input_dim, activation='sigmoid')(encoded)

autoencoder = keras.Model(input_layer, decoded)
autoencoder.compile(optimizer='adam', loss='binary_crossentropy')
autoencoder.fit(X_train, X_train, epochs=20, batch_size=256)

Notice that the model is trained with X_train as both the input AND the target, since the whole point is reconstructing the input itself.

Why the Bottleneck Forces Useful Learning

Because the compressed representation has far fewer dimensions than the original input, the network cannot simply copy the input through unchanged. It is forced to learn which features are genuinely most important for reconstruction, discarding redundant or less informative details.

Use Case: Dimensionality Reduction

The trained encoder alone can compress high-dimensional data into a much smaller, dense representation, similar in spirit to PCA but capable of learning non-linear relationships that PCA cannot capture.

Use Case: Anomaly Detection

An autoencoder trained only on normal data learns to reconstruct normal patterns well. When it encounters an anomalous input, reconstruction error tends to be much higher, since the model never learned to compress that unusual pattern effectively. This makes autoencoders useful for fraud detection and equipment failure detection.

reconstructions = autoencoder.predict(X_test)
reconstruction_error = np.mean(np.square(X_test - reconstructions), axis=1)

threshold = np.percentile(reconstruction_error, 95)
anomalies = X_test[reconstruction_error > threshold]

Use Case: Denoising

A denoising autoencoder is deliberately trained on noisy input paired with a clean target output, teaching it to remove noise, useful for cleaning up corrupted images or sensor data.

Common Interview Questions

How does an autoencoder learn without labelled data?

It uses the input data itself as the target, training the network to reconstruct its own input, which counts as a form of self-supervision rather than requiring separate labels.

Why can autoencoders be used for anomaly detection?

An autoencoder trained only on normal data reconstructs normal patterns well but struggles with anomalous ones it never learned to compress, so an unusually high reconstruction error signals a likely anomaly.