PolyU | Artificial Intelligence Concepts

This lecture introduces the fundamental definitions of Artificial Intelligence, distinguishing between Artificial Narrow Intelligence (ANI) and the theoretical Artificial General Intelligence (AGI). It also explores the field's historical origins, including the Turing Test’s role in operationalizing intelligence and the 1956 Dartmouth Conference that established Symbolic AI as a core discipline.

This lecture introduces problem-solving agents, which transition from simple reflex-based actions to goal-oriented planning using formal state-space representations. It further explores the infrastructure of search algorithms, contrasting the memory-efficient but cycle-prone Tree Search with the more robust Graph Search that utilizes an explored set to prevent redundant work.

This lecture introduces adversarial search and constraint satisfaction, focusing on how agents make rational decisions in competitive, multi-agent environments. Students will learn to apply algorithms like Minimax and Alpha-Beta pruning for strategic gaming, as well as techniques for solving constraint-based problems through variable assignment and backtracking.

This lecture explores Knowledge Representation and Reasoning (KRR), focusing on how AI systems use symbolic logic to model the world and perform explicit, verifiable inference. Students will learn about the transition from classical logical foundations like Propositional and First-Order Logic to modern approaches like Knowledge Graphs and Neuro-Symbolic AI.

This lesson explores the shift from deterministic computing to the probabilistic nature of Generative AI, where models generate creative outputs by sampling from high-dimensional probability distributions. Students will learn how to distinguish between aleatoric and epistemic uncertainty and understand why managing these statistical processes is essential for balancing AI creativity with factual reliability.

This lecture introduces the fundamentals of classical machine learning, distinguishing between supervised learning for predictive tasks and unsupervised learning for pattern discovery. Students will explore key algorithms like decision trees and K-means clustering while learning to implement them using the standardized Scikit-learn API.

This lecture explores the transition from symbolic, rule-based AI to connectionist models that utilize biologically-inspired, bottom-up learning. Students will learn how artificial neural networks mathematically abstract biological structures—such as dendrites and synapses—into inputs, weights, and summation functions to process data and identify patterns.

This lecture introduces the fundamentals of computer vision, distinguishing between digital image processing for signal enhancement and computer vision for semantic interpretation. It further explains why Convolutional Neural Networks (CNNs) are superior to Multi-Layer Perceptrons for visual tasks by utilizing local connectivity, weight sharing, and translation invariance to overcome the challenges of high-dimensional image data.

This lecture introduces sequence modeling by highlighting the importance of temporal order, autoregressive properties, and variable-length data processing. It further explores the evolution of these concepts from the state-based memory of Recurrent Neural Networks (RNNs) to the attention-driven architecture of Transformers.

This lecture explores the paradigm shift from task-specific models to unified Large Language Models (LLMs) that utilize a monolithic transformer architecture for diverse linguistic tasks. Students will learn about the LLM lifecycle, including pre-training, fine-tuning, and reinforcement learning from human feedback (RLHF), while examining the scaling laws that drive emergent model capabilities.

This lecture explores the transition from monolithic models to multi-layered Compound AI Systems, focusing on the orchestration of infrastructure, foundation models, and agentic workflows. It also examines the technical drivers of LLM hallucinations, explaining how the reliance on statistical token prediction over grounded truth creates a reliability gap in generative outputs.

This lecture explores the shift from performance-focused AI to a Responsible AI (RAI) framework, emphasizing the need to treat development as a constrained optimization problem to ensure safety and fairness. Students will learn to bridge the gap between controlled benchmark performance and real-world robustness by applying ethical guardrails and analyzing systemic risks.

This lab introduces the Google Colab environment, focusing on mounting Google Drive for persistent storage, managing Python packages with pip, and utilizing GPUs for efficient tensor computations. Students also explore essential data structures and interactive widgets to prepare for upcoming assignments in symbolic AI and deep learning.

This lab introduces Symbolic AI through the classic SHRDLU "Blocks World," where students learn to represent world states, implement symbolic physics, and use regex for natural language parsing. The assignment progresses from direct execution to automated planning, teaching students how to use recursive logic to resolve physical constraints and clear obstacles in a simulated environment.

This lab introduces the fundamentals of building a Gomoku AI by covering 2D grid representation, nested loops for board traversal, and direction vectors for win detection. Students will also learn to implement the Minimax algorithm with recursive thinking, heuristic scoring, and backtracking to enable the AI to evaluate future game states efficiently.

This lab focuses on optimizing Gomoku AI by implementing strategic heuristics, such as positional heatmaps, and performance enhancements like spatial locality and move ordering. These techniques reduce the search space and improve the efficiency of the Minimax algorithm, allowing the AI to search deeper and make more intelligent decisions.

This lab provides a comprehensive guide on structuring a technical assignment report for COMP5511, covering essential sections like the abstract, introduction, background theory, and methodology. Students will learn how to effectively document their work using Markdown, apply academic writing standards, and clearly explain complex AI concepts such as symbolic planning and adversarial search.

This lab introduces the transition from symbolic AI to machine learning by using the MNIST dataset to train a Support Vector Machine (SVM) classifier. Students learn to process image data using PyTorch tensors, normalize pixel values, and apply Scikit-Learn’s standardized workflow to flatten and classify handwritten digits.

This lab introduces Convolutional Neural Networks (CNNs) as a solution to the limitations of classic machine learning by preserving spatial information in images through filters, activation functions, and pooling. Students learn to implement these architectures in PyTorch using nn.Module and understand the iterative training process of forward passes, loss calculation, and backpropagation.

This lab explores the limitations of traditional machine learning models like SVMs when handling complex, high-dimensional data and demonstrates how deep CNNs overcome these challenges through GPU acceleration and batch processing. Students will learn to scale their models using PyTorch to efficiently process larger datasets like CIFAR-10, highlighting the necessity of parallel computing in modern deep learning.

This lab introduces the transition from computer vision to Natural Language Processing by exploring how language models generate text through iterative token prediction. Students will gain hands-on experience using the Hugging Face library to load, tokenize, and run instruction-tuned mini-LLMs like Qwen-0.6B on local hardware.

This lab introduces fine-tuning as an efficient way to adapt a pre-trained Qwen model to a specific persona using a small, conversationally formatted dataset. Students will learn to prepare data, apply chat templates, and use tokenization techniques to transform a standard assistant into an emoji-based communicator.

This lab explores how to overcome the limitations of Large Language Models by building AI Agents that combine a "Brain" (the LLM) with "Hands" (Python tools and APIs) to interact with the real world. Students learn to implement an iterative "Think-Act-Observe" loop, using system instructions and regex parsing to enable the AI to fetch and process live data.

This lab introduces the ReAct (Reasoning + Acting) pattern, which improves AI performance by forcing models to "think out loud" before executing tasks. Students will learn to overcome the limitations of stateless LLMs by implementing a scratchpad memory and an iterative loop that integrates external Wikipedia tool data as observations.

This mock exam provides a self-paced review and practice opportunity for the PolyU COMP5511 course to help students prepare for the upcoming assessment. It serves as a tool to test your understanding of key course concepts and evaluate your readiness for the final exam.