• can learn features directly from images
• typically consist of a multiple convolutional, non-linear activation, and pooling layers.
  • The primary purpose of these layers is to extract features
• The final feature maps are then transformed in a one-dimensional vector and sent to a feedforward neural network, which maps the feature representation to the desired output

The methods

• Convolution process for image classification
• Maxpooling
• Image Padding

Network architecture

Learning hierarchical features

• Early layers tend to learn low-level features such as edges and simple shapes
• deeper layers capture more complex and abstract features,
  • including object parts and high-level patterns.
• enables deep NNs to understand and represent data at multiple levels of abstraction.
• Using an SOM:

Receptive field

(A receptive field of a neuron with respect to the input)

• = is the area in the input that the particular neuron can “see”.
• As we move deeper
  • the receptive field (in terms of the input) increases in size
  • allowing deeper layers
  • to capture and understand
  • larger, more complex regions of the input image.

Extra interactive resources:

• Interactive site explaining CNNs
• Build your own NN

Some more NB concepts

• Backpropagation
• Dropout