Skew Logistic Distribution Applied as Activation Function in Artificial Neural Networks

In recent years, Artificial Neural Networks (ANNs) have stood out among machine learning algorithms in many applications, such as image and video pattern recognition. Activation functions play a crucial role in the operation of these algorithms, directly influencing the effectiveness of ANNs. The logistic (or sigmoid) function is often used as a standard activation function, but the existing literature lacks in-depth investigations into the potential of the Skew Logistic (SL) function. This work investigates the SL as an activation function in ANNs, exploring its ability to handle imbalanced data distributions. To achieve this, the function was implemented and evaluated on four binary classification datasets using accuracy, precision, recall, and F1-score. The results show that SL improved performance on the datasets selected for experiments 1 to 3, where an increase was observed in some performance metrics compared to the sigmoid function. And in experiment 4, competitive performance was obtained with softmax, using a multiclass version of SL. In particular, it was noted that it is possible to improve precision or recall, adjusting the asymmetry parameter ( $\lambda $ ). It is concluded that the SL function offers a viable and promising alternative to conventional activation functions.