Gen-CNN: a framework for the automatic generation of CNNs for image classification

Convolutional neural networks (CNNs) have become widely adopted for computer vision tasks. However, the vast amount of design choices and the complex interactions among their hyperparameters, which ultimately influence the model’s performance, impede their accessibility to users who are not experts in machine learning (ML). To address this challenge, we present AutoML as a solution, leveraging hyperparameter optimization (HPO) for effective parameter selection. Particularly good at handling non-convex, non-differentiable optimization tasks, genetic algorithms are easy to implement and parallelize, making them well suited for deep learning applications. In this context, we introduce Gen-CNN, an AutoML framework based on a genetic algorithm that generates CNN models for image classification. Our framework incorporates transfer learning and operates in a low-compute regime to accelerate the hyperparameter optimization phase. We test Gen-CNN on four datasets, including Sign Language Digits for convergence assessment and KVASIR-v2, ISIC-2019, and BreakHis for performance evaluation. Our results prove that Gen-CNN automatically generates CNN models with classification performance comparable to state-of-the-art custom models already published in the literature. Moreover, in the recommended testing regime for heuristic optimization techniques, we surpassed other HPO algorithms by achieving better mean categorical accuracy. Gen-CNN code is available at—omitted for anonymous review. © The Author(s), under exclusive licence to Springer-Verlag London Ltd., part of Springer Nature 2024.