Qu'est-ce que Tanh ?
Tanh, short for hyperbolic tangent, is a mathematical function defined as the ratio of the hyperbolic sine and hyperbolic cosine. It is expressed with the formula:
tanh(x) = sinh(x) / cosh(x) = (e^x – e^(-x)) / (e^x + e^(-x))
The Tanh function maps any real-valued number to a range between -1 and 1. This characteristic makes it particularly useful in various fields, especially in intelligence artificielle and apprentissage automatique, where it is commonly used as an fonction d'activation in réseaux neuronaux.
By squashing input values into a limited range, Tanh helps to normalize the outputs of neurons, ensuring that the data fed into subsequent layers remains manageable and conducive to learning. Compared to the sigmoid activation function, which only outputs values between 0 and 1, Tanh provides a symmetric output centered around zero, which can lead to faster convergence during training.
One of the main advantages of using Tanh is its ability to reduce the likelihood of the vanishing gradient problem, a common issue in apprentissage profond where gradients become too small for effective weight updates. However, Tanh can still face challenges such as saturation, where inputs that are too high or too low can lead to gradients near zero, slowing down learning.
En résumé, Tanh est une fonction mathématique essentielle dans le domaine des réseaux de neurones, offrant un moyen d'effectuer des transformations non linéaires tout en maintenant les sorties dans une plage gérable.