Course Contents

• What is learning, Introduction to Neural Networks, Backpropagation TA session: Introduction to Pytorch and Use case 1: Basic NN from scratch
• Introduction to CNN, CNN layers, insights into CNN TA Session: Use case 2: Construction of CNN from scratch
• Various architectures for CNN, CNN visualization, attention TA Session: Use case 3: Using pre-trained model, Use case 4: Image segmentation for the real world
• Auto-encoders, Image segmentation, Object Detection TA Session: Use case 5: Processing videos, Use Case 6: Real-time Object Detection
• RNN, LSTM, Intro to generative models, MLE, MAP TA Session: Use case 7: Setting up your Own LSTM, Use case 8: LSTM for image captioning
• Idea of GAN, and Overview before mid-semester TA Session: Use case 9: LSTM for Time series, Use case 10: Machine Translation, Use case 11: Image generation using GAN
• Contrastive learning, Language modeling, RNN based encoding and decoding, RNN variants: LSTM, GRU, Attention from decoder to encoder states, Application: POS, NER, translation, summarization, RNN cell enhancements: GRU, LSTM
• Transformer architecture, Positional embeddings, Attention structure variations, Masked language modeling, Generative pretraining, KV-cache, Applications: sentiment, POS, NER, Pretraining and fine-tuning
• The need for retrieval in generative models, Brief history of lexical retrieval engines, Sparse and dense vector spaces, Vector databases, nearest neighbor search
• Dense passage retrieval (single vector), Clustered indexes, small world networks, Multi- vector retrieval: ColBERT
• Sparsified representation for lexical indices: SPLADE, Training dense retrieval systems, Motivation for retrieval from passages, graphs, tables, Knowledge graphs (KG), KG representation, completion, Graph neural networks and, graph transformers
• (Passage) retrieval augmented reasoning and generation, REALM, Fusion in decoder, RAG, Unlimiformer, (KG) retrieval augmented reasoning and generation, EmbedKGQA, GraphRAG, Thinking-on-Graph, Paths-over-Graph
• Prompt and prefix tuning, Instruction Fine Tuning, RLHF, Chain of Thought, PEFT, LORA, QLORA, Reinforcement Learning and LLM; RLHF, Transfer Learning and 0- shot Learning
• Few-shot learning, Domain and language adaptation, Performance optimizations

System of Evaluation

A participant is awarded a grade based on his/her performance in examinations/assignments in every course registered by him/her. These grades are described by the letter
AA, AB, BB, etc. and have a numerical equivalent called grade points as given below: