Image Classification

Input: Image NxN pixels

Annotation require lots of expertise. People doing the annotation must be very knowledgeable.

Useful for real-time object detection

A good dataset is CIFAR-10.

Convolutional Network Architecture (CONVNET)

Various Layers of CONVNET 1D

Layers in Convolutional Network

• Input Layer
• Convolutional Layer
• Max Pooling
• Feed-forward Network
• Output layer
• Error

Backpropagtion can be used all the way, and can be used for weights of inputs as well.

Convolutional Layer

Here we're learning filters, which is a set of tied inputs that are learned. We learn these filters through back propagation.

Common representation: learned weights as a filter used to do a convolution on the input.

Max Pooling

This just takes the max of the outputs.

Various Layers of CONVNET 2D

Convolutional Layer

Matrix multiplication where the original matrix is multiplied by a filter, to create a result of the element-wise product and sum of the filter matrix and the original matrix.

Max-pooling layer

Addition layer (with no adjustable weights) can be introduced to further reduce dimensionality: the pooling layer. It's just downsampling using a max operation.