This paper presents a survey of activation functions (AFs) used in deep learning (DL) applications and highlights recent trends in their use. The authors compile a comprehensive list of AFs, emphasizing their applications in practical DL deployments and comparing them with state-of-the-art research results. The paper aims to assist in selecting the most suitable AF for specific applications. It is timely as it is the first to compare AF trends in practice against research findings in DL. Deep learning algorithms use AFs to perform complex computations between hidden and output layers. AFs are crucial for learning patterns in data and improving model performance. Common AFs include Sigmoid, Tanh, Softmax, Softsign, ReLU, and variants like Leaky ReLU, Parametric ReLU, and ELU. Each AF has unique properties, such as non-linearity, gradient behavior, and computational efficiency, which influence their suitability for different tasks. The paper discusses the evolution of AFs over time, highlighting their roles in various DL applications like image recognition, natural language processing, and speech recognition. It also addresses challenges such as vanishing and exploding gradients, which AFs help mitigate. The study emphasizes the importance of AFs in deep learning, noting that while newer AFs often outperform older ones, many established AFs like ReLU remain widely used due to their simplicity and effectiveness. The paper concludes that while there is ongoing research into new AFs, practical applications still rely on well-tested and proven functions. Future work could involve comparing these advanced AFs on standard datasets to assess their performance improvements. Overall, the paper provides a detailed overview of AFs in DL, their applications, and the current trends in their usage.This paper presents a survey of activation functions (AFs) used in deep learning (DL) applications and highlights recent trends in their use. The authors compile a comprehensive list of AFs, emphasizing their applications in practical DL deployments and comparing them with state-of-the-art research results. The paper aims to assist in selecting the most suitable AF for specific applications. It is timely as it is the first to compare AF trends in practice against research findings in DL. Deep learning algorithms use AFs to perform complex computations between hidden and output layers. AFs are crucial for learning patterns in data and improving model performance. Common AFs include Sigmoid, Tanh, Softmax, Softsign, ReLU, and variants like Leaky ReLU, Parametric ReLU, and ELU. Each AF has unique properties, such as non-linearity, gradient behavior, and computational efficiency, which influence their suitability for different tasks. The paper discusses the evolution of AFs over time, highlighting their roles in various DL applications like image recognition, natural language processing, and speech recognition. It also addresses challenges such as vanishing and exploding gradients, which AFs help mitigate. The study emphasizes the importance of AFs in deep learning, noting that while newer AFs often outperform older ones, many established AFs like ReLU remain widely used due to their simplicity and effectiveness. The paper concludes that while there is ongoing research into new AFs, practical applications still rely on well-tested and proven functions. Future work could involve comparing these advanced AFs on standard datasets to assess their performance improvements. Overall, the paper provides a detailed overview of AFs in DL, their applications, and the current trends in their usage.