https://cs231n.github.io/convolutional-networks/

Intro

1. nearby pixels often correlate most with nearby pixels (locality)
2. statistics of pixels relatively uniform across image (stationary statistics)
3. Identity of an object doesn’t depend on its location in the image
4. objects are made of parts

• The locality property of the visual world is represented in Conv Nets by receptive fields that are only connected to a small patch of the image
• The stationary statistsics property of the visual world is represented in Conv Nets by Learned features applied all over the images
• The translation invariance property of the visual world is represented in Conv Nets by receptive fields that are only connected to a small patch of the image and Max Pooling

Kernels

• think of it as a lens

Screen Shot 2024-02-04 at 10.03.34 PM.png

Conv Nets

• Feed forward: convolve input
• introduce non linearity
• pooling
• train Convolutional filters by backprop classification error