cse7321 viva questions

7321c_Q.txt

What is perceptron
Fullyconnect ante enti?
Shallow vs Deep Learning (Why Deep)
Issues with Deep Learning
Uses of Unsupervised methods
What are auto encoders
Loss functions for Regression, Classification
What are dropouts and their suggested range
Three key advantages of CNNs over NNs
Which paratemer influences the depth of feature map
What is image augumentation


How does RNN work?
Formula for Simple RNN?
What is CNN1D, how is it useful
Which parameters influence output dim of a feature map (kernel size, stride, padding, number of f-maps)
How stacking autoencoders work/What are stacked autorencoders
How autoencoders can be used as feature generators
How are weights initialized 
Range of learning rates and momentum (2 separate questions)
How MLP will help as opposed to perceptron
List all activations and range of each activation function
Advantage of Relu over Sigmoid
How is loss minimized - Gradient Descent 
Use of Gradient Descent
Difference between Contractive and Denoising autencoders
What is Local Connectivity, Spatial Arrangement, Parameter Sharing
What does pooling achieve (Local translational and rotational invariance)


Which is simpler model
	1 layer with 256 nodes or
	2 layers with 16 nodes each?
	How do you justify?
Batch Normalization ante enti?
What is regularization
Regularization Techniques
	EarlyStopping; WeightDecay; Dropouts; BatchNorm; L1; L2
What are embeddings?
Dimension of an embedding matrix



###########################################################
What are the assumption of Linear Programming
What is sensitivity analysis of shadow price
Pipeline of genetic algorithm
Difference between mutation crossover