Convolutional Neural Network (CNN) Cont.

(24th-September-2020)

Convolution + pooling is repeated (= Deep CNN) to acquire invariance to various deformations • Learnable - target: filter coefficients,

Method: Backpropagation

Pooling type

Pooling type - Average pooling, MAX pooling, L2 (Lp) pooling • Objective of pooling = Acquire invariance - Translate, rotate, scale multiples of local patterns - more than that? • Key to Object Recognition - Compatibility of discrimination and invariance