Comprehensive Natural Language Processing (NLP)

Course Outline

Introduction to Natural Language Processing

• History of Natural Language Processing and Generative AI
• What are Natural Language Processing models?
• Understanding generative models
• Contrasting generative and discriminative models

NLP Use Cases and Tasks

• NLP use cases in insurance, finance, and consulting
• Examples of NLP products

Tensorflow and Keras

• What is TensorFlow and Keras?
• Tensors and Python API
• TensorFlow Lite
• TFX (TensorFlow Extended)
• TensorFlow Toolkit Stack
• TensorBoard
• Core Keras data structures
• Components of a Keras model
• Creating neural networks in Keras
• The strengths and weaknesses of the functional API

Neural Networks and Deep Learning

• What is a neural network?
• Types of neural networks
• Deep Learning
• Navigating neural network layers
• Neurons
• Perceptrons and multi-layer perceptrons
• Neural network training

Traditional NLP Models: RNNs and LSTMs

• What are RNNs and LSTMs?
• Feedforward neural networks vs. RNNs
• Mathematical foundations of RNNs and LSTMs
• Training and inference on RNNs and LSTMs
• Limitations of RNNs and LSTMs

Text Representations

• The semantic aspect of word embeddings
• Word embeddings in NLP
• Word embeddings in transformers
• Word embedding similarity metrics

Transformers

• What is a transformer?
• Transformer use cases
• Encoders and decoders
• Self-attention mechanism
• Multi-head attention
• Tokenization
• Transformer architecture

Foundation Models

• What are foundation models or Large Language Models (LLMs)
• Multimodality of foundation models
• Model training techniques
• LLM capabilities vs. size
• Model quantization
• Options for accessing LLMs
• Context window and prompts
• Foundation model architecture
• Diffusion models and the diffusion process

Prompt Engineering and Retrieval-Augmented Generation (RAG)

• Prompt engineering basics
• Prompt types
• Prompt context
• Prompt pitfalls
• One-shot prompting
• Few-shot prompting
• Chain-of-Thought prompting
• Automated prompt engineering
• What is RAG?
• Vector indexing
• LlamaIndex indexes: List, Vector Store, Tree, Keyword
• LangChain: Model I/O, Retrieval, Chains, Memory, Agents

Pre-training and Fine-tuning Large Language Models

• Building the pre-trained LLM
• Benefits and challenges of pre-training
• Pre-training phases
• Fine-tuning over pre-training
• Data preparation and processing
• Data augmentation: text, image, and signal data
• When to use fine-tuning
• Domain adaptation
• Multi-task learning

Evaluating LLMs

• Diversity Metrics
• Likelihood
• Perplexity
• Inception Score
• BLEU
• ROUGE
• Human evaluation
• LLM-on-LLM evaluation

Security and Privacy

• What is AI cybersecurity?
• Threats and challenges in AI cybersecurity
• Adversarial attacks
• Model inversion and extraction
• Data poisoning
• Robustness techniques
• Differential privacy
• Federated learning
• Best practices in secure AI development

Responsible AI

• What is an AI system?
• Principles of AI Ethics
• Fairness
• Accountability
• Transparency
• Privacy and autonomy
• Reliability
• Explainable AI (XAI)
• Model-Agnostic Visual Analytics (MAVA)
• Human-AI Collaborated Evaluation (HACE)
• AI ethics in practice
• Regulatory compliance in AI systems

Lab Exercises

• Learning the Colab Jupyter Notebook Environment
• Basics of Tensorflow and Keras
• Build a neural network to classify insurance fraud
• Dive into RNNs and LSTMs
• Dive into word embeddings
• Implement transformer models for finance tasks
• Compare LLM models for insurance tasks
• Prompt engineering and RAG for insurance question-answering
• Fine-tune an LLM for insurance topic classification
• Evaluating LLMs
• Nemo Guardrails for finance tasks
• Design an LLM audit process