可视化 ReLU 激活函数输出。
激活函数是神经网络非线性的来源。
| Item | Formula / Note | Source / Standard |
|---|---|---|
| ReLU definition | ReLU returns the maximum of the input and zero so the negative side is flat | Nair & Hinton 2010, ICML |
| LeakyReLU definition | LeakyReLU keeps a small slope on the negative side to avoid vanishing gradients | Maas et al. 2013, ICML |
| Gradient behaviour | ReLU is non-differentiable at zero and its zero negative gradient can kill units | Glorot et al. 2011, AISTATS |