延伸理論:三個研究方向Three directions beyond the core theorems

IDL115 · CH03 · KP16 · 43/44
激活、架構、數學工具擴展收斂分析的範圍。 / Theory extends across activations, architectures and tools.

激活函數Activations

二次、ReLU、Leaky ReLU:各自需要對應條件。 / Each activation has its own assumptions.

模型與資料Architectures and data

ResNet、結構化資料與不同初始化。 / Residual networks and structured data.

分析工具Analytical tools

臨界點、分類損失曲面、代數幾何、寬網誤差界。 / Critical points, algebraic geometry and error bounds.