📄 中文摘要
深度神经网络的训练过程，尽管看似随机，实则展现出一定的稳定性。通过随机初始化权重并结合梯度下降法进行迭代更新，这些模型能够有效地收敛到训练损失的全局最小值。关键在于，优化过程中的步长始终保持在初始点附近，从而确保了学习路径的良好形状。
随机梯度下降优化超参数深度 ReLU 网络
📄 English Summary
Stochastic Gradient Descent Optimizes Overparameterized Deep ReLU Networks
Deep neural network training, despite appearing random, exhibits stability. Through random weight initialization and gradient descent iteration, these models effectively converge to the global minimum of training loss.