An Evolutionary Multi-objective Optimization of Vision Transformer

This paper presents an evolutionary multi-objective optimization of vision transformer architecture and the effect of two genetic crossover operators, two-point and uniform crossover, on the quality of the solutions. The proposed method employs the multi-objective genetic algorithm (NSGA-II) and uses a block-based encoding for visual transformer models. By adopting configurable blocks, the encoding method allows the search for various ViT architectures. The paper comprehensively analyzes the effect of genetic operators on the quality of the solution, providing practical insights for future research. The experimental results, which are reported on the Cifar-10 dataset, demonstrate a trade-off between model size and accuracy and are compared to the state-of-the-art models, providing solid evidence of the findings.