Accelerating deep learning model development—towards scalable automated architecture generation for optimal model design

In deep learning, CNNs are powerful tools for various applications. Manual architecture selection is time-consuming, limiting dataset exploration. This paper introduces a novel algorithm for CNN architecture, revolutionizing the way models are designed and evaluated. By providing the number of convolutions and essential hyperparameters, researchers can unleash the algorithm’s prowess to generate, train, and compare model configurations. This algorithm excels in identifying the most efficacious model named OptiNet for its outstanding performance and facilitates a data-driven method for customizing optimal architectures. Comprehensive experiments across diverse datasets have underscored the algorithm’s ability to strike a delicate balance between model complexity and generalization. The algorithm creates N = 2^c−1 models, where N represents the number of all possible models using c convolutions. This new method can help reduce the difficulty of manually creating architectures. The algorithm empowers researchers to make informed decisions, achieve competitive results, and streamline model development for cutting-edge deep learning.