目录

• Feature maps
• Why not Linear
  • 335k or 1.3MB
  • em...
• Receptive Field
• Fully connnected
• Partial connected
• Locally connected
• Rethink Linear layer
• Fully VS Lovally
• Weight sharing
• Why call Convolution？
• 2D Convolution
• Convolution in Computer Vision
• CNN on feature maps

Feature maps

• 单通道

• rgb三通道

• rgb三通道合成

• 数字2的卷积成像图

Why not Linear

• 4 Layers: [784, 256, 256, 256, 10]

335k or 1.3MB

em...

• 486 PC + AT&T DSP32C

  • 256KB
  • 66Mhz

• Batch X

• Gradient Cache

• etc.

Receptive Field

Fully connnected

Partial connected

Locally connected

Rethink Linear layer

Fully VS Lovally

Weight sharing

• 三阶张量的卷积

• 6 Layers

  • ~60k parameters

• 4 layers, 335k

Why call Convolution？

2D Convolution

y(t)=x(t)∗h(t)=∫∞−∞x(τ)h(t−τ)dτ$$y(t)=x(t)\ast h(t)={\int }_{-\mathrm{\infty }}^{\mathrm{\infty }}x(\tau )h(t-\tau )d\tau$$

Convolution in Computer Vision

• 模糊化

• 边缘检测

CNN on feature maps