He Initialization, nomeado em homenagem a its creator Kaiming He, is a technique used to initialize the weights of redes neurais. This method is particularly effective for layers using the Rectified Linear Unit (ReLU) funções de ativação, which are common in aprendizado profundo modelos.
O objetivo principal da inicialização de pesos é evitar o problema do gradiente que desaparece ou explode, which can hinder the training of deep networks. When weights are initialized too small, the network may not learn effectively (vanishing gradients), and when they are too large, the gradients can become excessively large, leading to instability (exploding gradients).
He Initialization addresses these issues by setting the initial weights to random values drawn from a distribuição normal with a mean of 0 and a variance of 2/n, where n is the number of input units in the layer. This scaling factor helps to maintain a balanced signal flow through the layers of the network, facilitating effective learning.
This method is especially beneficial for deep networks, as it allows for larger learning rates without the risk of gradient-related problems. By ensuring that the weights are appropriately scaled, He Initialization contributes to faster convergence during training and can lead to better desempenho geral da rede neural.
Em resumo, He Initialization é uma técnica fundamental no campo do aprendizado profundo que melhora o treinamento de redes neurais, especialmente aquelas que usam ativações ReLU, fornecendo uma estratégia robusta para a inicialização de pesos.