L'Adaptive Moment Estimation, communément appelé Adam, est une méthode avancée algorithme d'optimisation widely used in l'entraînement de modèles d'apprentissage automatique, particularly in apprentissage profond. Introduced in 2014 by D.P. Kingma and J.B. Ba, Adam combines the ideas of two other popular des techniques d'optimisation: Élan and RMSProp.
L'avantage principal d'Adam est sa capacité à ajuster de manière adaptative le taux d'apprentissage for each parameter based on the first and second moments of the gradients. This means that during training, Adam tracks both the average of the gradients (first moment) and the variance of the gradients (second moment). By doing so, it can make more informed updates to the model parameters, allowing for faster convergence and improved performance.
Adam utilise deux hyperparameters, beta1 and beta2, which control the decay rates of the moving averages of the gradients and their squared values, respectively. Typically, beta1 is set to 0.9 and beta2 to 0.999. Another crucial parameter is epsilon, a small constant added to prevent division by zero during updates.
One of the key features of Adam is its robustness to noisy gradients and its efficiency in handling sparse data. This makes it particularly suitable for applications in traitement du langage naturel, image recognition, and other fields where gradient noise is prevalent. Overall, Adam has become a go-to choice for many practitioners due to its simplicity and effectiveness in optimizing complex models.