Neural Networks and Deep Learning (ELEDSE7B) — DU B. Sc. (Hons) Electronics Previous Year Question Papers

Total Time : 3 Hours 1. 2. 3. Attempt 5 questions in all. (a) How is Deep Learning different from conventional Machine Leaming? (b) What is a Hidden Layer in an Artificial Neural Network (ANN)? What is its main function? (c) Explain the basic architecture of a RNN. How is it different from the Back Propagation Neural Network? (d) Given the dataset : Label (Rain) = [1,1, 0, 0, 0] Write the basic steps required to implement a Neural Network model to classify whether it may rain or not. (e) What are hypexparameters in deep learning models? Identify the types of hypexparameters used during training and elaborate on any one of them. (I) What are the key characteristics of ResNet? Where is it used? (a) Explain the convolutional operation on the following grey scale image matrix using a simple 2x2 filter and a stride s=2. What will be the final feature map? (b) Without activation functions, neural networks act as linear models. Justify this statement. Provide the mathematical fomulation of any three activation functions that help overcome this limitation (9) (a) Backpropagation is widely used for training deep neural networks. Describe how it works and present the mathematical update rules used in Gradient Descent (6) (b) Discuss how Radial Basis Neural Networks (RBNNs) use REF kernels to map data into a higher-dimensional feature space. Explain the architecture and show how this results in nonlinear hyperplanes for curve fitting (6) (c) RNNs extend BPNNs by introducing memory through recurrent connections. Explain the RNN architecture, contrast it with BPNNs, and describe how the BPTT algorithm enables learning from sequential data (6) (a) Explain the PCA technique in detail. Discuss all mathematical operations involv6driuch as standardization, covariance matrix computation, eigenvalues, and eigenvectors (6) (b) Write the mathematical connection between PCA eigenvectors and the left/right singular vectors obtained using SVD. Use a diagram or block representation (6) ( c) True target y=3 .0. Three model predictions (from repeated training/runs) are [2.5, 3.5, 3..0]. Compute the squared bias, variance of predictions, and mean squared error (MSE). Verify decomposition MSE = bias2 + variance (6) (a) Describe the purpose of the pooling layer in a Convolutional Neural Network (CNN). (c) A convolutional layer has a single weight w=0.5. The gradient of the loss with respect to thisweightisg=;;=-0.8.Theleamingrateisn=0.I Compute the updated weight using basic Stochastic Gradient Descent (SGD). Also, compute the updated weight using SGD with momentum (P=0.9) assuming initial velocity vo=0. (6) (a) What is meant by "perception" in self-driving cars? Discuss the types of sensors used for environmental perception in self-driving cars. (6) (b) What are the key differences between retrieval-based and generative chatbots? Explain their pros, cons, and practical applications. (6) (c) What role does prompt engineering play in using LLMs effectively? A Large Language