Deep learning Flashcards
What are the step of progressive encoding?
1) Input
2) Segmentation
3) Completion
4)Spatiality
5)Modeling
6)Semantics
Basic idea of CNN
local receptive field
shared weights
rectified linear units and pooling
2D Convolution formula
O[m,n] = f[m,n]*g[m,n]=
= sum_u(-inf, inf)(sum_v(-inf,inf)(f[u,v]g[m-u,n-v])
What are the main task achieved by convolution
Filtering and downsampling
Formula of a neuron in a convolutional layer
u_i_m = f(b + sum_l(1,L)_m(1,M)(w[l,m]a[j-1+l,k-1+m]))
f can be tanh for example
formula of output size of convolutional layer
sx =(Ix+2zx-fx)/Lx +1
Two ways to process RGB images in convolutional layers
1) Using three parallel lines of processing
2)Using a 3D filter to include include all three channel in the computation of each pixel of the feature map
Complete series of layer for the processing of an image and what they do
Input
Conv layer: filtering
ReLU layer: makes everything positive
POOL layer: smoothing
Main kinds of pooling
Max and average
Softmax classifier
It uses normalized exponential activation with many neurons as the number of classes
The cost function is the negative log probability of the correct label (cross-entropy regime)
What is the main difference between traditional ANN and deep learning
In traditional ANN we apply learning techniques to well-know quantities. Usually we start from raw data, we extract relevant features by preprocessing and then we train a model
In deep learning we do not know exaxtly what are the features that determine the task. End to end the model execute the task directly on the raw data