Lecture 07 - Autoencoders

Lecture07

Autoencoders

Available Functions

• demo(): Train autoencoder for MNIST handwritten digits
• display_latent_space(encoder, X_flat, Y_label): Display 2D latent space visualization
• display_reconstruction(model, image): Display original vs reconstructed image

Usage

using MachineLearningCourse
Lecture07.demo()

using MachineLearningCourse
X_flat, Y_label, autoencoder, encoder = Lecture07.demo(display=false)
i = 1 # set image index
Lecture07.display_reconstruction(autoencoder, X_flat[:, i])

using MachineLearningCourse
X_flat, Y_label, autoencoder, encoder = Lecture07.demo(display=false)
Lecture07.display_latent_space(encoder, X_flat, Y_label)

create_autoencoder(input_dim, hidden_dim, latent_dim)

Create symmetric autoencoder with encoder-decoder architecture.

Arguments

• input_dim::Int: Input dimension (e.g., 784 for MNIST)
• hidden_dim::Int: Hidden layer dimension
• latent_dim::Int: Latent space dimension (bottleneck)

Returns

• Flux.Chain: Autoencoder model with architecture:
  • Encoder: input → hidden (ReLU) → latent (linear)
  • Decoder: latent → hidden (ReLU) → output (sigmoid)

The model learns to compress input to latent space and reconstruct it.

demo(; train_size=5000, epochs=50, hidden_dim=128, latent_dim=2, display=true)

Complete MNIST autoencoder demo with 2D visualization.

Arguments

• train_size::Int: Number of training samples (default: 5000)
• epochs::Int: Number of training epochs (default: 50)
• hidden_dim::Int: Hidden layer dimension (default: 128)
• latent_dim::Int: Latent space dimension (default: 2)
• display::Bool: Whether to display latent space plot (default: true)

Returns

• Tuple: (X_test, model, encoder, plot)
  • X_test::Array{Float32, 2}: Flattened test images (784 × 1000)
  • Y_test::Int: Labels
  • autoencoder: Trained autoencoder model
  • encoder: Encoder part of the autoencoder

display_latent_space(encoder, X, Y)

Display 2D latent space visualization colored by digit class.

Arguments

• encoder: Trained encoder model
• X::Array{Float32, 2}: Flattened images (784 × n)
• Y::Vector{Int}: Digit labels (0-9)

Returns

• Plot object

display_reconstruction(model, image)

Display original vs reconstructed image comparison.

Arguments

• model: Trained autoencoder model
• image::Array{Float32, 1}: Flattened image vector (784 elements)

Returns

• Tuple: (plot, error)

extract_encoder(model)

Extract encoder part from trained autoencoder.

Arguments

• model: Trained autoencoder model

Returns

• Flux.Chain: Encoder portion (input → latent space mapping)

Useful for visualizing latent representations or transfer learning.

train!(model, X_train; epochs=50, η=0.001f0, batch_size=256)

Train autoencoder using reconstruction loss (MSE).

Arguments

• model: Autoencoder model to train
• X_train: Training data (features × samples), values in [0,1]
• epochs::Int: Number of training epochs (default: 50)
• η::Float32: Adam optimizer learning rate (default: 0.001f0)
• batch_size::Int: Mini-batch size (default: 256)

Returns

• Vector{Float32}: Training losses per epoch

Trains the autoencoder to minimize MSE between input and reconstruction.