AKTU Deep Learning (KCS078) Question Paper 2024-25

Note: Attempt all Sections. In case of any missing data; choose suitably.

Section A

Q1. Attempt all questions in brief · 2 × 10 = 20 marks

1. Discuss the differences between a perceptron and a support vector 1 K 1 , machine in terms of decision boundary formation. K
2. Discuss how the choice of a loss function affects model performance for 1 K 1 , a classification problem.
3. Trace the historical development of deep learning that enabled deep 2 K , networks to outperform traditional machine learning models.
4. How does weight sharing in CNNs improve efficiency and 2 K 1 , performance?
5. Explain the impact of weight initialization on the performance of deep networks.
6. Compare performance of autoencoders with traditional dimensionality reduction techniques.
7. Provide examples of optimization algorithms used for deep networks. 4 K
8. How do techniques address the challenges of stochastic optimization in 4 K deep learning?
9. Describe the architecture and training of WaveNet for audio generation.
10. Discuss the role of Word2Vec in natural language processing.

Section B

Q2. Attempt any three of the following · 10 × 3 = 30 marks

1. Derive the gradient of the logistic regression loss function with respect 1 K 1 , to its parameters. How does regularization impact its optimization process?
2. Define VC dimension and discuss its implications for understanding the 2 K , generalization capabilities of deep networks.
3. Explain the role of manifold learning in dimensionality reduction. How does it compare to linear methods?
4. Explain the concept of generalization in neural networks. How do techniques like dropout and weight regularization improve K generalization performance?
5. Discuss the pipeline of scene understanding using deep learning models. How does this pipeline handle challenges like occlusion and varying lighting conditions?

Section C

Q3

Attempt any one part of the following · 10 × 1 = 10 marks

1. Prove that neural networks can act as universal function approximators. 1 K 1 , Discuss any constraints or assumptions involved in this theorem. K
2. Discuss the relationship between the choice of activation function in a 1 K 1 , shallow neural network and its ability to model non-linear relationships.

Q4

Attempt any one part of the following · 10 × 1 = 10 marks

1. Analyze the trade-offs between increasing network depth and the risk of 2 K 1 , vanishing/exploding gradients. What role does skip connections in ResNet play in addressing these issues?
2. How does semi-supervised learning differ from supervised and 2 K 1 , unsupervised learning? Discuss its application in training deep networks K with limited labeled data.

Q5

Attempt any one part of the following · 10 × 1 = 10 marks

1. Explain the concept of metric learning in the context of dimensionality reduction. How does it differ from traditional techniques like PCA
2. Analyze the architectural differences between AlexNet, VGG, Inception, and ResNet. How do these architectures address the vanishing gradient problem?

Q6

Attempt any one part of the following · 10 × 1 = 10 marks

1. Discuss the implications of non-convex optimization in designing deep learning models. How do saddle points influence the training process? K
2. Provide a detailed overview of computational neuroscience principles applied in designing artificial neural networks.

Q7

Attempt any one part of the following · 10 × 1 = 10 marks

1. Provide a detailed overview of deep learning-based techniques for audio detection. How does WaveNet outperform traditional methods?
2. Describe the process of joint detection and captioning in image captioning systems. How does attention enhance this process?

Short on time? See the most repeated AKTU Deep Learning questions across all years.

Other years

More AKTU Deep Learning papers on LearnAIFast.

This paper is reproduced for study purposes only. Copyright belongs to AKTU. Source: the question paper bank of ABES Institute of Technology, an AKTU-affiliated college. If you own the copyright and want it removed, email learnwithskillnest@gmail.com and we will take it down.