Lectures

You can download the lectures here. We will try to upload lectures prior to their corresponding classes.

• 

  Course Introduction
  tl;dr: An introduction to the course syllabus, timeline and background.
  [slides]
  
• 

  Regular expressions & morphology
  tl;dr: Terminologies, regular expressions, morphology, Porter stemmer
  [slides]
  
• 

  Minimum Edit Distance & Statistical Language Modeling
  tl;dr: Edit distance table, backtracking, probabilistic language modeling, n-grams, smoothing
  [slides (Edit Distance)] [slides (N-grams)]
  
• 

  Advanced Smoothing Techniques and Evaluation of Language Models
  tl;dr: Backoff and interpolation; Good-Turing; Kneser-Ney; Shanon Game; perplexity and entropy
  [slides]
  

  Supplementary Material:

  • Language Models: Advanced Smoothing & Evaluation
  • Perplexity and Entropy: Jursfsky and Martin Page 59
• 

  POS Tagging and Hidden Markov Model
  tl;dr: Intro to POS tagging, rule-based methods, Markov chain, Intro to HMM, Forward algorithm, Viterbi algorithm
  [slides (POS Tagging)] [slides (HMM)]
  
• 

  Hidden Markov Model
  tl;dr: Evaluation, learning and decoding
  [scribe]
  

  Supplementary Material:

  • Jurafsky and Martin
• 

  Parsing
  tl;dr: Introduction to syntactic parsing, constituency vs dependency parsing, CFG
  [slides (Statistical Parsing)] [slides (CFG)]
  
• 

  Lexical Semantics
  tl;dr: Word relations, WordNet and word similarity
  [slides]
  
• 

  Distributional Semantics
  tl;dr: Word similarity, TF-TDF, distributional similarity
  [slides]
  
• 

  Word Representation
  tl;dr: Various methods for representing words as vectors – count-based methods, prediction-based methods (Word2vec, fastText)
  [slides]
  

  Supplementary Material:

  • Textbook: Chapter 3 – Introduction to Large Language Models, Tanmoy Chakraborty
• 

  Word Representation-II and Neural Language Models
  tl;dr: Co-occurrence matrices and GloVe; word vector properties; historical word embeddings; bias detection, fixed-window neural language models; transition to recurrent architectures (RNNs, LSTMs)
  [slides (Word Representation)] [slides (Neural Language Models)]
  

  Supplementary Material:

  • Textbook: Chapter 3 – Introduction to Large Language Models, Tanmoy Chakraborty
  • Textbook: Chapter 5 – Introduction to Large Language Models, Tanmoy Chakraborty
• 

  Neural Language Models
  tl;dr: RNNs, training RNNs, backpropagation through time
  [slides]
  

  Supplementary Material:

  • Textbook: Chapter 5 – Introduction to Large Language Models, Tanmoy Chakraborty

  • Reading
  • Vanishing Gradient proof
  • Blog post
• 

  RNNs and Seq-to-Seq Attention
  tl;dr: vanishing gradient problem; LSTMs and GRUs; bidirectional and multi-layer RNNs; neural machine translation; seq2seq models; beam search decoding, top-k sampling
  [slides (RNNs)] [slides (Seq-to-Seq Attention)]
  
• 

  Seq-to-Seq Attention and Transformers
  tl;dr: seq2seq attention; variants of attention; introduction to positional encoding
  [slides (Seq-to-Seq Attention)] [slides (Transformers)]
  

  Supplementary Material:

  • The Illustrated Transformer
  • Attention Is All You Need
  • Visualizing A Neural Machine Translation Model (Mechanics of Seq2seq Models With Attention)
  • Sequence to Sequence Learning with Neural Networks
  • Neural Machine Translation by Jointly Learning to Align and Translate
• 

  Transformers - Positional Encodings
  tl;dr: sinusoidal positional encoding, rotary positional encoding, batch and layer normalization
  [slides]
  

  Supplementary Material

  • RoFormer: Enhanced Transformer with Rotary Position Embedding
  • The Illustrated Transformer
  • Attention Is All You Need
• 

  Pretraining Strategies
  tl;dr: ELMo architecture, pretraining of BERT, masked language modeling and next sentence prediction
  [slides]
  

  Supplementary Material

  • Deep contextualized word representations
  • BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
• 

  Pretraining and Tokenization Strategies
  tl;dr: Pre-Training BART, T5, GPT and LLaMa family, causal masking, sub-word tokenization, Byte-Pair Encoding, wordPiece tokenization, unigram language model tokenization
  [slides (Pretraining Strategies)] [slides (Tokenization Strategies)]
  

  Supplementary Material

  • BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
  • Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
  • Neural Machine Translation of Rare Words with Subword Units
  • Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation
• 

  Prompt-Based Learning
  tl;dr: zero-shot, few-shot and in-context Learning, prompt sensitivity, prefix tuning, Chain-of-Thought, Tree-of-Thought, Graph-of-Thought, POSIX, instruction tuning
  [slides]
  

  Supplementary Material

  • Language models are few-shot learners
  • Chain-of-thought
  • POSIX: A Prompt Sensitivity Index For Large Language Models
  • Graph of Thoughts: Solving Elaborate Problems with Large Language Models
  • Tree of Thoughts: Deliberate Problem Solving with Large Language Models
• 

  Instruction Tuning
  tl;dr: Discussion on instruction tuning of LLMs – data collection, loss function, and properties of tuned models.
  [slides]
  

  Supplementary Material

  • Finetuned Language Models are Zero-Shot Learners
  • Orca: Progressive Learning from Complex Explanation Traces of GPT-4
  • WizardLM: Empowering Large Language Models to Follow Complex Instructions
  • Self-Instruct: Aligning Language Models with Self-Generated Instructions
• 

  Alignment of Language Models
  tl;dr: LLM training stages and alignment; limits of instruction tuning; RLHF with human or AI feedback; reward model using Bradley-Terry preferences; REINFORCE and gradient tricks; Q-function and advantage estimation; PPO for stable policy optimization.
  [slides]
  

  Supplementary Material:

  • Training language models to follow instructions with human feedback
  • Reinforcement Learning for Language Models
  • Constitutional AI: Harmlessness from AI Feedback
• 

  Alignment of Language Models - DPO
  tl;dr: DPO objective; Parametric vs non-parametric setup; KL regularization; Bias in DPO; Online vs offline DPO; Key takeaways
  [slides] [scribe]
  

  Supplementary Material:

  • Direct Preference Optimization: Your Language Model is Secretly a Reward Model
• 

  Guest Lecture - Retrieval-Augmented Generation (Dinesh Raghu)
  tl;dr: Intro to RAG; Closed-book vs open-book LLMs; Hallucinations & retriever failure; PEFT & LoRA; RAFT and domain adaptation issues; Context & paraphrase augmentation; Tool-calling with LLMs
  [slides]
  

  Supplementary Material:

  • RAGAS: Automated Evaluation of Retrieval Augmented Generation
  • Atlas: Few-shot Learning with Retrieval Augmented Language Models
  • Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
  • NEAREST NEIGHBOR MACHINE TRANSLATION
  • A Survey on Retrieval-Augmented Text Generation
  • Augmented Language Models: a Survey
  • REALM: Retrieval-Augmented Language Model Pre-Training
• 

  Tool Augmentation with LLMs
  tl;dr: Intro to tool-augmented LLMs; Toolformer for API usage; Limits of current tool-calling; SyReLM for symbolic solver coordination; Adapter and LoRA finetuning; DaSLaM decomposer-solver model; Reward functions for decomposition; Future of modular and tool-augmented LLMs
  [slides]
  

  Supplementary Material:

  • Toolformer
  • SyReLM
  • Survey
  • Optional Read
• 

  Knowledge Editing
  tl;dr: Intro to knowledge editing; Problems with full finetuning; Knowledge triplet updates; Conditions for reliable, generalized, localized edits; KE method; GRACE cache-based updates; ROME mid-layer memory editing; Evaluation on factual correction and retention
  [slides]
  

  Supplementary Material:

  • Locating and Editing Factual Associations in GPT
  • Editing Factual Knowledge in Language Models
  • Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors
  • Knowledge Editing for Large Language Models: A Survey
• 

  Responsible LLMs & Conclusion
  tl;dr: Responsible LLMs – Explainability, fairness, robustness, safety; Bias in LLMs – visibility, sources, impact; Bias mitigation via adversarial triggers and in-context learning; Final course summary; Future directions; NLP applications
  [slides]
  

  Supplementary Material:

  • A course on Computational Ethics for NLP
  • Blog
  • Cognitive bias in NLP