Neural networks employing ReLU or GELU activation functions are more difficult to interpret than networks using SwiGLU activation functions