AI academy - Fix, Fill, and Create: Reconstructing Missing Data and Synthesising New Ones with Autoencoders

This workshop develops an applied understanding of autoencoders through the problem of recovering useful information from incomplete or corrupted data.

The keynote begins with the encoder-decoder idea: a model learns to map high-dimensional input into a smaller latent representation and then reconstruct the original signal. Participants examine bottlenecks, reconstruction loss, undercomplete representations, convolutional autoencoders, and the relationship between representation learning and dimensionality reduction.

The case study uses a visual dataset that is progressively degraded with noise, occlusion, and missing regions. Participants first train a basic reconstruction model, then convert it into a denoising autoencoder and test how well it can recover structured information. Reconstruction error is introduced as a practical signal for anomaly detection, while latent representations are visualised to examine similarity and structure.

In the hands-on lab, participants build a convolutional autoencoder in Keras/TensorFlow, generate corrupted training inputs, compare reconstruction losses, and inspect the learned latent space. A guided extension demonstrates latent interpolation and introduces the idea behind variational autoencoders as a route from reconstruction toward controlled generation.

Morning

Keynote Lecture: Learning to Compress and Reconstruct

Topics covered through the keynote:

• Encoder-decoder architectures and latent representations
• Bottlenecks, reconstruction losses, and information preservation
• Dense and convolutional autoencoders
• Denoising autoencoders and corrupted inputs
• Reconstruction error for anomaly detection
• Latent spaces, similarity, and interpolation
• From autoencoders toward generative latent-variable models

Hands-on Lab: Fixing and Reconstructing Visual Data

• Build a convolutional autoencoder in Keras/TensorFlow
• Create noisy and masked versions of image inputs
• Train and evaluate reconstruction quality
• Inspect latent representations and reconstruction error
• Explore interpolation between learned representations
  
  

Afternoon

Group Case Study: Fix, Fill, or Create

Each group selects a practical reconstruction or generation objective and defines:

• the type of corruption, missingness, or variation to model
• the autoencoder design and reconstruction objective
• how success should be measured
• how latent representations will be inspected
• risks associated with reconstructed or synthetic outputs
  
  

Group Presentations and Model Critique

Short presentations (approximately 5 minutes per group)
Discussion of fidelity, anomaly detection, and responsible synthetic data use

By the end of the course, participants will be able to:

• explain the encoder-decoder structure and purpose of an autoencoder
• build and train a convolutional autoencoder with Keras/TensorFlow
• use autoencoders for denoising and reconstruction of corrupted inputs
• interpret reconstruction loss and reconstruction error
• inspect and reason about learned latent representations
• describe how latent spaces can support interpolation and generative modelling
• evaluate practical and ethical limitations of reconstructed and synthetic data
  
  

One-day intensive workshop combining:

• case-study keynote lecture
• hands-on autoencoder implementation in Jupyter notebooks
• guided denoising and reconstruction experiments
• group-based reconstruction or generation challenge
• presentations and model critique