CONVNETS-classification, detection, segmentation,self-driving cars, pose recognition, medical image diagnosis etc..

fully connected layer vs convolution layer

• fully connected layer - 32323 image → stretch to 3072 *1
• convolution layer - 32 * 32* 3 image → preserve spatial structure

convolution layer

convNet- a sequence of convolution layers, interspersed with activation functions

ex) CONV,RELU → CONV,ReLU → CONV, ReLU

→ each grid is the value that maximizes the activation function of a certain neuron

(what type of image is the neuron looking for?)

closer look at spatial dimensions

• Output size: NN image convloution with FF filter → (N-F) / stride +1

• In practice : zero padding → (N+2*P-F) / stride +1

    → It is common to see CONV layers with S=1, F*F, zero padding with (F-1)/2
  

Pooling layer

-makes the representations smaller, manageable