ИСТИНА

Deep Neural Networks (DNNs) have proven effective in unsupervised novelty detection tasks, including model-independent searches in High-Energy Physics. This study investigates the impact of hidden layer activation functions on the accuracy of novelty detection. Our findings highlight that both the accuracy and training stability of the final model are significantly influenced by the choice of activation function. Rational activations, a recent advancement in learnable activation functions, offer the potential for high expressiveness with a minimal number of learnable parameters. By incorporating these activations into one-class DNNs, we aim to enhance their accuracy and stability.