Supervised deep learning

Question 1

What is one distinctive feature of deep neural networks?

Answer 1

learning takes place directly on raw data, without any pre-processing or feature selection. The latter are discovered by the network during learning.

Question 2

What are Convolutional Neural Networks composed of?

Answer 2

at least one convolutional layer, in which neurons are not fully connected with the preceding layer but have local receptive fields
The convolutional layer is followed by a pooling layer
In the deepest part of the network there is at least one “standard” hidden layer, that is a layer that is fully connected with the preceding one

Question 3

What is the role of the pooling layer?

Answer 3

to reduce the dimensionality of the image (max pooling operation). This reduces the number of parameters, controlling overfitting and promoting invariance (e.g. traslation)

Question 4

What do the neurons in the convolutional layer have?

Answer 4

Each hidden neuron of the convolutional layer has a local receptive field (filter) which encodes a specific feature