Kernekoncepter
Quality Diversity evolution with MEliTA algorithm enhances multimodal creative tasks by promoting coherence and diversity in generated artefacts.
Resumé
This content discusses the application of the MEliTA algorithm, a variation of MAP-Elites, for handling multimodal creative tasks. It introduces an innovative approach to improve text-to-image mappings within the solution space. The paper explores the use case of generating text descriptions and cover images for hypothetical video games, showcasing how MEliTA can lead to fitter and more diverse outcomes compared to traditional MAP-Elites algorithms.
Directory:
Abstract:
Recent advances in language-based generative models enable orchestration of multiple generators into one system.
Introduction:
Traditional struggles in evaluating artefacts in creative domains due to lack of universal metrics.
Data Extraction:
"The most interesting development in the field of deep learning is the training and release of multimodal models."
"MAP-Elites partitions the solution space into a multi-dimensional grid."
Use Case: Generating Text & Visuals for Game Titles:
Utilizes GPT-2 model for text generation and Stable Diffusion for image generation.
Experimental Protocol:
Evaluates performance metrics such as mean fitness, max fitness, coverage, and QD score over 2000 parent selections.
Results:
MEliTA outperforms MAP-Elites in producing fitter individuals but with lower coverage.
Discussion:
Highlights limitations in variation operators and potential improvements for future work.
Conclusion:
Concludes that MEliTA enhances quality and diversity in multimodal creative tasks.
Statistik
"The most interesting development in the field of deep learning is the training and release of multimodal models."
"MAP-Elites partitions the solution space into a multi-dimensional grid."