Softmax de Gumbel
Le Gumbel Softmax is a method utilisé en apprentissage automatique to enable differentiable sampling from categorical distributions. This is particularly useful in scenarios where we want to optimize models that require discrete decisions, such as in apprentissage par renforcement or modèles génératifs.
Traditional categorical sampling is not differentiable, which poses a challenge when using gradient-based optimization methods. The Gumbel Softmax overcomes this limitation by introducing a continuous relaxation of the categorical distribution. It does this by using the Gumbel distribution, which allows us to sample from the categorical distribution in a way that is differentiable.
In practical terms, the Gumbel Softmax works by adding Gumbel noise to the logits of the categorical distribution, followed by applying the softmax function. The result is a soft approximation of a one-hot vector that represents the sampled category. As the temperature parameter in the softmax function approaches zero, the output becomes increasingly similar to a encodage one-hot, effectively simulating a discrete choice while remaining fully differentiable.
This method is advantageous in various applications, including neural networks for natural language processing, where decisions are often categorical, and in variational autoencoders, where it allows for efficient training while still sampling discrete variables. Overall, the Gumbel Softmax is a powerful tool for bridging the gap between discrete and continuous optimisation en apprentissage automatique.