Fundamentals of Deep Learning and Generative AI Training

Introduction to Neural Networks and Deep Learning

• What is an Artificial Neural Network?
• Types of Neural Networks
• Machine Learning with Neural Networks
• Deep Learning
• Navigating Neural Networks Layers
• Positional Types of Layers
• The Network and the Model
• Model Properties
• A Bit of Terminology
• Data Pre-processing
• How Does My Network Know Which Problem I Want It to Solve?
• A Neuron
• The Artificial Neuron
• The Perceptron
• The Perceptron Symbol
• A Breakthrough in Neural Networks Design
• Perceptrons and MLPs
• A Basic Neural Network Example
• Popular Activation Functions
• Supervised Model Training
• Measuring the Error with the Loss (Cost) Function
• Mini-batches and Epochs
• Neural Network Training Steps
• Applying Efficiencies with Autodiff …
• Neural Network Libraries and Frameworks

Neural Network Concepts and Terminology

• Why We Need Terminology …
• Features and Targets
• Observations (Examples)
• Notation for Observations
• Data Structures: Tensors
• Continuous and Categorical Features
• Continuous Features
• Categorical Features
• Feature Types Visually
• Feature Importance
• Supervised and Unsupervised Machine Learning
• Self-Supervised Learning
• Common Distance Metrics
• The Euclidean Distance
• Visualizing Data on the X-Y Plane
• What is a Model?
• Model Life-Cycles
• Model Parameters and Hyperparameters
• The Train/Validate/Test Machine Learning Triad
• Training/Validation/Test Data Split Ratios
• Data Splitting Considerations
• Cross-Validation Technique
• Test Data Leakage
• Bias-Variance (Underfitting vs Overfitting) Trade-off
• Bias and Variance Visually
• Model Underfitting vs Model Overfitting Visually
• Ways to Balance Off the Bias-Variance Ratio
• Training Error vs Validation Error Diagram
• Loss (Cost) Functions
• Loss Function Properties
• Mean Squared Error (MSE)
• Mean Absolute Error (MAE)
• (Categorical) Cross Entropy Loss
• The Cross Entropy Loss Visually
• The Chain Rule in Calculus
• The Chain Rule in Neural Networks
• Gradient Descent in Neural Networks (1/2)
• Gradient Descend Visually
• Gradient Descent in Neural Networks (2/2)
• An Annotated Example of Gradient Calculation
• The softmax Function
• Coding Softmax
• Model Accuracy in Classification Tasks
• Confusion Matrix
• The Binary Classification Confusion Matrix
• Multi-class Classification Confusion Matrix Example
• Feature Engineering
• Data Scaling and Normalization
• The Data Normalization Tooling
• Regularization
• A Hands-On Exercise
• Mathematical Formulations …
• Dimensionality Reduction
• Online Machine Learning Glossaries

TensorFlow Introduction

• What is TensorFlow?
• The TensorFlow Logo
• Tensors and Python API
• Python TensorFlow Interfaces Diagram
• PyTorch
• GPUs and TPUs
• Google Colab
• Data Tools
• TensorFlow Variants
• TensorFlow Core API
• TensorFlow Lite
• TFX (TensorFlow Extended)
• A TFX Pipeline Example
• XLA Optimization
• TensorFlow Toolkit Stack
• Keras
• TensorBoard

Introduction to Keras

• What is Keras?
• Keras 3.0
• Core Keras Data Structures
• Layers in Keras
• The Dense Layer
• Defining the Layer Activation Function
• Models in Keras
• Components of a Keras Model
• Creating Neural Networks in Keras
• The Sequential Model
• A Sequential Model Code Example
• The Strengths and Weaknesses of Sequential Models
• The Functional API
• A Functional API Example
• The Strengths and Weaknesses of the Functional API
• Making New Layers and Models via Subclassing
• A Layer Subclassing Example
• A Model Subclassing Example
• The Strengths and Weaknesses of Subclassing

Introduction to CNNs

• Convolutional Neural Networks (CNNs)
• Kernels and Convolutions
• A Convolution Mathematically
• A Convolution Visually
• A Quiz
• Kernels and Feature Maps
• Feature Maps in CNNs
• CNN Efficiencies
• Feature Maps Visually
• The Stride Hyperparameter
• The CNN Architecture
• The Conv2D Class
• A Quiz
• An Example of Pooling Layer
• Finally, Putting it All Together
• Summary

Introduction to RNNs

• Recurrent Neural Networks (RNNs)
• How Do RNNs Do It?
• Feedforward Neural Networks vs RNNs
• Mathematical Formulations
• (Simplified) RNN Visual Representations
• A More Accurate RNN Diagram
• Sampling the Data
• Problems with RNNs
• LSTM and GRU Networks
• Problems with LSTM and GRU Networks
• RNNs as a Precursor of Generative AI

Embeddings

• Embeddings …
• Understanding the Embeddings Visually
• Dimensionality
• The Semantic Aspect of Embeddings
• Word Embeddings in NLP
• Embeddings in Transformers
• Cosine Similarity

Introduction to Generative AI

• The Age of Digital Assistants …
• What is Generative AI?
• Applications
• What Are Natural Language Models?
• The Probabilistic Language Model
• Training a Language Model to Predict the Next Word
• Generative AI, the Pre-Cursor Technologies
• RNN Limitations
• Wait, there is More …
• Transformers
• The Problem Domain
• LLMs
• Multimodality of LLMs
• Infographic of Multimodality Tasks
• Generative Foundation Models
• Inferring Movie from Emoji
• Fine-Tuning and Transfer Learning
• Transfer Learning in Computer Vision
• The Transfer Learning Diagram
• Can I Have My Very Own Model?
• The Age of Digital Assistants … Transformed …
• The Training Datasets
• The Training Techniques
• Hugging Face
• The Evolutionary Tree of LLMs
• The LLM Capabilities vs LLM Size (in Parameters)
• Does the Model Size Matter?
• Inference Accuracy vs LLM Size
• OpenAI GPT Models
• ChatGPT
• The Microsoft 365 Copilot Ecosystem
• The LLaMA Family of LLMs
• LLaMA 2
• The AI-Powered Chatbots
• Options for Accessing LLMs
• Cloud Hosting
• Prompt Engineering
• Context Window and Prompts
• Zero- and Few-Shot Prompting
• Understanding Model Sizes
• Physical Model Sizes
• Quantization
• Generative Adversarial Networks
• Generator and Discriminator Networks
• A High-Level GAN Diagram
• The Above Generator’s Sample Output
• The Diffusion Models, Names, and “Competition”
• The Core Diffusion Modeling Idea
• The Diffusion Process
• AI Alignment
• Ethical AI

Introduction to Transformers

• What is a Transformer?
• Transformer Use Cases
• Transformers, Encoders, and Decoders
• Recurrent Neural Networks
• Why Transformers?
• The Transformer Evolution Path
• A Short Summary of the Transformer Inner Workings
• N-Grams
• Tokenization
• Two Types of Weights
• (Self-)Attention (1 of 3)
• Multi-Head Attention
• The Encoder
• The Decoder
• The Head-First Approach …
• The Transformer Model Architecture
• Model Training
• The Overall Translation (Inference) Process
• Positional Encoding
• The Encoding Part (a Big Picture)
• The Decoder Attention Units
• Cross-Attention
• The Decoder Part (a Big Picture)
• The Attentions Weights Matrix

Lab Exercises

• Lab 1. Learning the Colab Jupyter Notebook Environment
• Lab 2. Neural Network Playground Web App
• Lab 3. Multi-layer Perceptron Classifier
• Lab 4. Vectors and Matrix Operations
• Lab 5. Understanding the Gradient Descent Algorithm
• Lab 6. Understanding Regularization
• Lab 7. TensorFlow Basics
• Lab 8. Using Keras for Image Classification
• Lab 9. Using CNNs for Image Classification
• Lab 10. Understanding RNNs
• Lab 11. Word2vec Pre-Trained Embeddings
• Lab 12. Hello, Generative AI!
• Lab 13. Using OpenAI API
• Lab 14. NLP and NLU with Transformers