SUK · Third Year B.Tech. Computer Science and Engineering (Artificial Intelligence and Data Science) · NEP 2020
Deep Learning
AI&DS3201 · Semester VI · 3 credits · ISE 10 + mid-semester 30 + ESE 60
Unit 1: Neural Networks and Deep Learning
6 hoursAI, ML and deep learning and a brief history; why deep learning; neural network basics and data representation; gradient-based optimisation and the anatomy of a network.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- Representation learning
- Feed-forward network architectures
- Gradient descent variants
- Backpropagation
Unit 2: TensorFlow, Keras and Hyperparameters
6 hoursInstalling TensorFlow; computation graphs, tensors, rank and tensor maths, NumPy and tensors; Keras models, layers and preprocessing; hyperparameters: learning rate, iterations, hidden layers and units, activation, momentum and mini-batch size; overfitting, underfitting and regularisation.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- Deep learning frameworks: TensorFlow, Keras and PyTorch
- Hyperparameters in deep learning
- Momentum and Nesterov accelerated gradient descent
- Regularisation and dropout
Unit 3: Convolutional Neural Networks
7 hoursThe convolution and max-pooling operations; training a convnet from scratch on a small dataset; using a pre-trained convnet; visualising what convnets learn.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- Convolution and feature maps
- Padding, stride and pooling
- Training a CNN
- Transfer learning
- Visualising CNNs: guided backpropagation and Deep Dream
Unit 4: Sequence Models
6 hoursOne-hot encoding and word embeddings; recurrent layers in Keras; LSTM and GRU layers; advanced use of recurrent networks.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- Word embeddings: Word2Vec and GloVe
- Recurrent neural networks
- LSTM
- GRU
- Bidirectional and deep RNNs
Unit 5: Advanced Deep Learning Best Practices
4 hoursThe Keras functional API beyond sequential models; inspecting and monitoring models with Keras callbacks and TensorBoard; getting the most out of models.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- Keras functional API, callbacks and TensorBoard
- Debugging strategies for deep networks
Unit 6: Generative Deep Learning
7 hoursText generation with LSTM, DeepDream and neural style transfer; generating images with variational autoencoders; generative adversarial networks.
Examined in the 60-mark end-semester paper, and in the mid-semester exam if it is taught by then
- Encoder-decoder sequence models
- Autoencoders and VAEs
- Generative adversarial networks
Unit contents are summarised from the official SUK syllabus . AiCanCode.org is not affiliated with Shivaji University, Kolhapur.