Deep Learning — DU B.Sc. (Hons) Computer Science Previous Year Question Papers

Parts of a question should be attempted together. (f) What is data augmentation? List any two areas where it is applied. (4) (g) Define weight sparsity and name the regularization method used to achieve (3) it. Why is this method particularly useful when dealing with high-dimensional datasets containing redundant features? Q2 (a) Explain the working of the backpropagation algorithm. Describe the chain (8) rule of calculus and its role in backpropagation. (b) What is dropout? Explain how dropout improves the generalization ability (4) of deep learning models. Discuss the advantages and limitations of dropout in deep neural networks. (c) Identify and explain two key limitations of n-gram models that are (3) effectively mitigated using word embeddings in • neural network architectures. Q3 (a) In the context of CNN, an input grayscale image of dimension (5x5) is (9) Perform the convolution operation with the following kernel of dimension With stride = 1, and no padding: (b) On the feature map obtained in Q3 a) above, apply the ReLU activation (3) function and then perform max pooling with a 2x2window and stride = 1. (c) A weight has an initial value of 0.8, and the gradient of the loss with respect (3) to this weight is 0.5. Given a learning rate of 0.05, find the new value of the weight after one step of gradient descent. Q4 (a) Discuss the paradigm shift in speech recognition that occurred when Deep (6) Neural Networks (DNNs) replaced Gaussian Mixture Models (GMMs). What specific limitations of GMM-HMM systems did DNNs successfully address? (b) Describe the importance of activation functions in neural networks. (5) Write the mathematical expressions for the following activation functions: (iii) Linear Also, calculate the output of each activation function when the input value is xis4. (c) An autoencoder is designed for image compression with the following (4) Input Layer: 1,024 units (representing a 32×32pixel image) Bottleneck Layer: 128 units Output Layer: 1,024 units Assuming the network is fully connected and there are no bias terms, calculate the total number of trainable parameters. QS (a) Describe the architecture of an LSTM cell with the help of a clear diagram. (6) Explain the roles of the input gate, forget gate, output gate, and the cell state. (b) Explain why saddle points and flat regions (plateaus) slow down the (5) optimization process during the training of neural networks. (c) Consider a simple recurrent neural network (RNN) where the hidden state (4) is updated using the rule:h, = tanh (0.5 x+ + 0.4 ht-1) Here, Xt represents the input at time step t, and his the hidden state at time t. Given the input sequence X=[1,2] and the initial hidden state ho =0 compute the hidden states, h1 andhz. Q6 (a) Show that a single-layer perceptron cannot realize the NOR function. (6) Construct a two-layer neural network with step activation that implements