Programme Content
MATHS FOR GENAI (5 SESSIONS)
Linear Algebra & Probability Basics: Vectors, matrices, matrix multiplication, dot product, idea of rank, SVD (high level); basic probability definitions.
Probability & Optimization: Bayes theorem, idea of a distribution; first and second order conditions, idea of gradient descent.
Introduction to ML: ML intro and terminology; linear regression; classification with logistic regression.
Evaluation & Unsupervised Learning: Overfitting and regularization; evaluation measures; high-level ideas of K-Means and PCA.
Neural Networks: Neural networks and backpropagation.
NATURAL LANGUAGE PROCESSING (4 SESSIONS)
Introduction to NLP: Stemming, Porter stemmer, lemmatization, edit distance.
Statistical Language Models: Language modelling, n-grams, smoothing, evaluation, perplexity.
POS Tagging & Parsing: POS tagging with HMM, Viterbi, evaluation; constituency vs dependency parsing, CFG, PCFG, CKY algorithm.
Semantics: Lexical & Distributional: Lexical similarity: words and senses; distributional similarity, vector space model, PMI, MI, TF-IDF.
NEURAL LANGUAGE MODELS (2 SESSIONS)
Word Representation: One-hot encoding, Word2Vec, GloVe, evaluation.
CNN & RNN++: CNNs for text; RNN, LSTM, GRU.
TRANSFORMER ARCHITECTURE (2 SESSIONS)
Transformer I: Seq2Seq, Beam Search & Attention: Seq2Seq, beam search, attention mechanism.
Transformer II: Encoder-Decoder: Transformer encoder and decoder.
LM PRETRAINING & FINE-TUNING (2 SESSIONS)
Transformer III: Pretraining Strategies: Pretraining strategies for effective domain adaptation.
Fine-Tuning Strategies: Fine-tuning strategies for task-specific performance.
INSTRUCTION TUNING, PREFERENCE TUNING & PROMPTING (4 SESSIONS)
IFT & Alignment: I: SFT and instruction tuning.
Prompt Engineering: Prompt engineering, LangChain.
Alignment: II: Value and policy optimization, classical reward model.
RLHF: RLHF with the TRL framework.
AUGMENTED LLM (1 SESSION)
RAG & Tool Augmentation: Methods to improve an LLM's ability to solve complex problems; Toolformer.
AGENTIC AI (3 SESSIONS)
Agentic AI: I: Foundations of LLM agents: planning, reasoning, tool use, memory.
Agentic AI: II: Multi-agent systems, orchestration frameworks.
Agentic AI: III: Agent evaluation, safety & deployment.
GENERATIVE AI FOR VISION (3 SESSIONS)
Vision LM: I: CNNs for image classification and segmentation, Vision Transformers, CLIP.
Vision LM: II: BLIP, LLaVA, Masked Autoencoder (MAE), Segment Anything Model (SAM).
Vision LM: III: SAM extensions, object detection basics, open-vocabulary object detection (OVOD).
ADVANCED TOPICS (2 SESSIONS)
Advanced Topic I: Small Language Models: Design of SLMs: pruning, distillation, and quantization.
Advanced Topic II: Responsible LLM: Bias and fairness, hallucination, safety and alignment, privacy, evaluation and governance of LLMs.
