A Comparative Study of Modern Activation Functions on Multi-Label CNNs to Predict Genres Based on Movie Posters

Abstract

Categorization of images based on their visuals into various genres has a crucial role in the recommendation system. However, multilabel classification poses significant challenges due to the complexity of assigning multiple labels to each instance. This study contributes to the understanding of how activation functions influence the efficiency and accuracy of multilabel CNNs and provides practical insights for selecting appropriate functions in movie poster classification tasks. This investigation focused on identifying the activation function that provided the fastest convergence, highest accuracy, and lowest computational cost or training time. The results show that the Leaky ReLU activation function achieved the fastest convergence and highest training accuracy with an top accuracy of 99.7% and GELU demonstrated the highest validation accuracy at 91.5% across the training iteration. Softplus showed convergence characteristics at epoch 14 while other in 30. The computational cost analysis revealed that ReLU was computationally efficient with training time of 1896 seconds. Overall, the Leaky ReLU activation function is identified as the most effective for multilabel CNNs, balancing convergence speed, accuracy, and computational cost.