*changes may apply
Introduction
- Gemini, ChatGPT and other impressive artificial models.
- Human and Artificial Intelligence, Turing test, brain, neurons, synapses
- Brief history of AI
- Terminology: Machine Learning (ML), Supervised vs. unsupervised learning, Neural Networks (NN), Deep Learning (DL), NLP, LLMs, Reinforcement Learning (RL).
- Typical applications of AI. How to join the revolution?
Before AI - A quick review of some classical models
- Basic statistics
- Linear regression
- PCA
- K-means
- Decision Trees
- Imbalanced Class
- Relevant python libraries: NumPy, Matplotlib, Pandas
Deep Learning
- Main models and methods. Training, testing and validation.
- Using Python Notebook and Colab.
- Main AI training tools: Keras and TensorFlow by Google, Pytorch.
- Hardware and software infrastructure: GPUs, Cuda, Dockers.
- Relevant python libraries: NumPy, Matplotlib, Pandas.
- Your first Neural Network model: Detecting patterns in tabular data.
Computer Vision: CNNs and Vision Transformers (ViT)
- Visual tasks: classification, detection, and segmentation.
- Why is Machine Vision difficult?
- NN building blocks and layers: single neuron, convolutions, pooling, fully connected layers, normalization, activation functions, loss.
- Learning the model weights: the back-propagation algorithm.
- NN architectures: feed-forward, recurrent, encoder-decoder, Siamese.
- Practical open-source vision NNs: Resnet, Yolo, Clip, SAM.
- Vision Transformers (ViT)
- Native Multimodality
- Computer Vision mini project: detecting objects in images
Natural Language Processing (NLP) using Transformers
- Large Language Models: capabilities and limitations, and the shift towards Small Training Large Language Models (LLMs), Attention mechanism and the Transformer architecture.
- Language Models (SLMs) for local execution (e.g., Llama, Gemma, Phi)
- Using LLM services in your application: Google Vertex AI, Gemini APIs, and prompt engineering best practices
- Parameter-Efficient Fine-Tuning (PEFT): LoRA and efficient adaptation techniques.
- Retrieval-Augmented Generation (RAG): From basic RAG to advanced techniques (Hybrid Search, Re-ranking, GraphRAG).
- Fine-tuning and Aligning LLMs: Instruction tuning, RLHF, and Direct Preference Optimization (DPO).
- Parameter-Efficient Fine-Tuning (PEFT): LoRA, QLoRA, and integration with Hugging Face’s ecosystem.
- NLP mini project: Building a business application using LLM APIs and a RAG architecture.
Deep Reinforcement Learning (RL)
- Introduction: reinforcement learning vs. non-interactive supervised learning.
- Methods: Q-Learning and Deep Q-Network.
- Advanced Methods (Overview): Proximal Policy Optimization (PPO) and Actor-Critic models.
- Applications: control and robotics, automated driving, finance, and the role of RL in GenAI (RLHF and Reasoning models).
- Reinforcement learning mini project: gaming.
Other applications and trends
- Time series forecasting in healthcare.
- Recommendation systems.
- Cyber security and authentication.
- Generative AI Architectures: Diffusion Models and Flow-matching
- Multimodal Generative AI: image, video and audio generation from text
Real-life AI projects in the industry
- Overview of Hugging Face ecosystem: pretrained models, datasets, and Spaces.
- Choosing an open source to start with.
- Handling data: collecting, filtering, cleaning, augmenting, preprocessing.
- Synthetic Data Generation: rare events, privacy, compliance. Synthetic Data Vault (SDV), Datagen (vision), GPT-based augmentation.
- Validation, testing and and Observability: Measuring model quality, LLM-as-a-Judge, and evaluation frameworks (e.g., Ragas, LangSmith)
- Using TensorBoard: Visualizing the training process, controlling convergence and overfit.
- Experimenting architectures, loss functions and hyper-parameters.
- Improving model speed and efficiency: optimization, models search, pruning, distillation, and Quantization (e.g., GGUF, EXL2)
- Agentic Workflows and Multi-Agent Systems: Moving beyond basic agents using frameworks like LangGraph and CrewAI. Algorithmic benefits: tool-calling, memory, and reasoning loops (ReAct).
- Hands-on: Build a basic research agent with memory and planning.
- Deploying your AI models for smartphones and edge devices and NPUs with TF Lite
- Deploying your AI models on Google Cloud Platform for scale and stability
- AI development lifecycle with Google Cloud tools
Responsible AI
-
- Personal, social, and economic implications
- Who is responsible? legal aspects
- Being fair and reducing biases
- Data privacy, compliance, and security
- Explainable AI
- AI Safety in Practice: Red Teaming, handling jailbreaks, and implementing Guardrails
- Dealing with deep fake
- Who owns the data, the models, and the code libraries?
- Dangers and opportunities for mankind
- Towards synergy of mankind and AI
*Changes may be applied in the Syllabus