Diversity in Generative Machine Learning to Enhance Creative Applications.

PhD thesis by Sebastian Berns.

Queen Mary University of London, July 2024.

Contents

  1. Introduction
  2. Background
    1. Generative Modelling in Deep Learning
    2. Generative Models as Creative Systems
    3. Use of Generative Models with Evolutionary Algorithms
    4. Evaluation of Generative Models
    5. Measuring Output Diversity
    6. Human Perception of Similarity
  3. Artistic and Creative Uses of Generative Models
    1. Introduction
    2. Automating Generative Deep Learning for Artistic Purposes
    3. Discussion
  4. Limitations of Conventional Generative Modelling
    1. Introduction
    2. Artefact Generation via Latent Space Search
    3. Methodology and Setup
    4. Experiments
    5. Results
    6. Discussion
  5. Increasing the Output Diversity of Generative Models
    1. Introduction
    2. Mode Balancing
    3. The Vendi Score
    4. Diversity Weights
    5. Proof-Of-Concept Study on Hand-Written Digits
    6. Discussion
  6. Similarity Estimation for the Evaluation of Diversity
    1. Introduction
    2. Methodology
    3. Study 1: Human vs. Computational Similarity Evaluation
    4. Study 2: Interpretation of Similarity Dimensions
    5. Discussion
  7. Related Work
    1. Data Biases in Machine Learning
    2. De-biasing Generative Models
  8. Conclusions
    1. Future Work