Mumbai University · Final Year Computer Engineering · Rev-2019 'C' Scheme
Deep Learning
CSDC8011 · Department Optional Course 5 · Semester VIII · 3 credits · Internal tests 20 (average of two) + end-semester 80
End-semester paper (3 hours): Six questions of 20 marks; you answer four. Each module's share of the paper follows its lecture hours.
One of the Department Optional Course 5 options. In this paper Q1 and Q6 span every module, and Q2 to Q5 cover the other modules.
Module 1: Fundamentals of Neural Networks
4 hoursBiological neurons; McCulloch–Pitts; perceptron and delta learning; multilayer perceptron for separable and non-separable classes; deep networks and their history.
Examined in the 80-mark end-semester paper (about 10% of it, by lecture hours); the class tests cover what has been taught by then
- Perceptron and feed-forward networks
- Deep learning vs machine learning
- History, advantages and challenges of deep learning
Module 2: Training, Optimisation and Regularisation
10 hoursFeed-forward DNNs; tanh, logistic, softmax, ReLU and leaky ReLU; squared-error and cross-entropy loss; backpropagation; GD, SGD, mini-batch, momentum, Nesterov, AdaGrad, Adam and RMSProp; bias–variance; L1, L2, dropout, weight decay, batch norm, early stopping and augmentation.
Examined in the 80-mark end-semester paper (about 26% of it, by lecture hours); the class tests cover what has been taught by then
- Activation functions in deep networks
- Loss functions
- Forward and backpropagation
- Gradient descent and vanishing and exploding gradients
- Optimisers: momentum, Nesterov, AdaGrad, RMSProp and Adam
- Regularisation: dropout, batch normalisation and early stopping
Module 3: Autoencoders
6 hoursLinear, undercomplete and overcomplete autoencoders; regularisation; denoising, sparse and contractive autoencoders; image compression.
Examined in the 80-mark end-semester paper (about 15% of it, by lecture hours); the class tests cover what has been taught by then
- Autoencoders
- Denoising, sparse and contractive autoencoders
Module 4: Convolutional Neural Networks
7 hoursConvolution, padding and stride; input, output and filter sizes; convolution and pooling layers; weight sharing; LeNet and AlexNet.
Examined in the 80-mark end-semester paper (about 18% of it, by lecture hours); the class tests cover what has been taught by then
- Convolution, padding and strides
- Pooling and fully connected layers
- CNN architecture
- LeNet and AlexNet
Module 5: Recurrent Neural Networks
8 hoursSequence learning; unfolding graphs; RNNs and bidirectional RNNs; BPTT and truncated BPTT; vanishing and exploding gradients; LSTM and GRU.
Examined in the 80-mark end-semester paper (about 21% of it, by lecture hours); the class tests cover what has been taught by then
- Recurrent neural networks
- Bidirectional and deep RNNs
- Backpropagation through time
- Long-term dependencies, LSTM and GRU
Module 6: Recent Trends
4 hoursGAN architecture; image generation and deepfakes.
Examined in the 80-mark end-semester paper (about 10% of it, by lecture hours); the class tests cover what has been taught by then
- Generative adversarial networks
- Types and applications of GANs
Module contents are summarised from the official Mumbai University syllabus . AiCanCode.org is not affiliated with the University of Mumbai.