Class 12 13 Flashcards
Structure of CNNs?
Grid like structure, 1D grid time series 2d grid images 3d grid videos.
The main idea is to apply a convolution operation to the input
The main steps of CNNs? And describe them
PES Valla PeS
Sparse Interactions: Unlike dense connections, in sparse connections not all the input and output units are connected with each other. This helps the network to have smaller weights, and reduce the memory requirements. It is like that because it depends on the kernel.
Parameter Sharing: The weights are shared in CNNS wihich help to improve the generalization. We just ned to keep the filter in the memory, the filter is applied on each of the input positions.
Computational complexity of sharing the weights or not is the same since it will be the same multiplications or additions. What really matters is the memory requirement because with shared weights, you only need to store the kernel.
Equivariance to trasnlation: f(g(x)) = g(f(x)), if g iis the translation, then CNN s are equivariaance to translation
Variable sized inputs: The size of the input wont affect the CNN, if the size increases than the feature map will be affected and it will also be increased.
Pooling(invariance to small translations)
Strided convolution: It reduces the size of the feature map by applying downsampling on the feature maps.
Increases the receptive field
what is spatial pooling and global pooling?
Spatial pooiling is pooling which gives CNN a small invariance to translation, global pooling is applied when there is a need of fixed size input, for example after cNN there is a fully connected layer which needs to have a fixes size network, therefore it is important to use global pooling
what is padding and why do we use that
We use padding when we dont want to lose the information which comes from the borders of the network
Unshared convolution and feature map?
Unshared convolution is the same but the weights are not shared across the network,
after applying the fillter to the network, we obtain a feature map
What is precision recall and ROC curve?
Precision: correctly detected fractions
Recall: true events detected fractions
Precision recall curve
fscore: 2precisionrecall / precision + recall
Hyperparameter selections?
The depth, the width of the network, learning rate.. ONLY THE LEARNING RATE IS TUNED IN THE TRAINING THE OTHERS ARE TUNED IN VALIDATION
Automataic hyperparameter selectins:
Select a range for hyperparameters,Grid search:test many pre defined options, Random search:usually just some of the hyper parameters make slight differences
what is a graph convolutional neural network?
Grahps are irregular grid structures where not the position but the connections of the vertices is important.
Graphs have different number of vertices, each vertex have a label I(v) and a vectorial attribute a(v) which has d vectorial attributes
What makes a graph hard to learn?
there are mainly two things what make a graph hard to learn
1) Isomorphism
2)Subgrahps
Problem 1: how to represent graphs into fixed sized vectors?
Problem 2: how to avoid explosion of thenumber of parameters
How do we define convolutions on graphs ?
Convolutions are normally in time and space domain, in order to use them in graphs we have to convert it in to frequency domain by using fourier transformation.
It means that we have to multiply pointwise in the frrequency domain
