DBATU · Final Year B.Tech Information Technology · Non-NEP
Deep Learning
BTITOE704C · Elective II · Semester VII · 3 credits · Continuous assessment 20 + mid-semester 20 + end-semester 60
One of the Elective II options.
Unit 1: Foundations
The history and success stories of deep learning; the McCulloch–Pitts neuron and threshold logic; perceptrons and their learning algorithm; multilayer perceptrons and their representation power; sigmoid neurons, gradient descent and feedforward networks.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- History and success stories of deep learning
- McCulloch–Pitts neuron and perceptron learning
- Multilayer perceptron and backpropagation
Unit 2: Training Feedforward Networks
Backpropagation; gradient descent, momentum, Nesterov, stochastic GD, AdaGrad, RMSProp and Adam; eigenvalues and eigenvectors, eigendecomposition, PCA and SVD.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- Gradient descent and optimisers
- Eigendecomposition, PCA and SVD
Unit 3: Autoencoders and Regularisation
Autoencoders and their relation to PCA; regularised, denoising, sparse and contractive autoencoders; the bias–variance trade-off, L2 regularisation, early stopping, data augmentation, parameter sharing, noise injection, ensembles and dropout.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- Undercomplete autoencoders
- Sparse and denoising autoencoders
- Contractive autoencoders
- Regularisation: dropout, L1 and L2
- Dataset augmentation, noise injection and ensembles
Unit 4: Word Vectors and Convolutional Networks
Learning vector representations of words; convolutional networks: LeNet, AlexNet, ZF-Net, VGGNet, GoogLeNet and ResNet; visualising CNNs, guided backpropagation, Deep Dream, Deep Art and fooling CNNs.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- Learning word vector representations
- CNN architecture
- AlexNet
- CNN architectures: ZF-Net, VGGNet, GoogLeNet and ResNet
- Visualising and fooling CNNs: guided backpropagation and Deep Dream
Unit 5: Recurrent Networks and Attention
Recurrent networks; backpropagation through time and its truncation; vanishing and exploding gradients; GRUs and LSTMs; encoder–decoder models and attention, including attention over images.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- Recurrent neural networks
- Backpropagation through time, vanishing and exploding gradients
- LSTM networks
- Encoder-decoder architectures
- Attention mechanism
Unit contents are summarised from the official DBATU syllabus . AiCanCode.org is not affiliated with Dr. Babasaheb Ambedkar Technological University (DBATU).