For the Technically Curious

This section is designed for faculty, researchers, and technical staff interested in understanding the broader taxonomy of AI techniques and how Generative AI fits within it. 

Defining Artificial Intelligence 

Core Definition: Artificial Intelligence (AI) is a field of computer science focused on creating systems that can perform tasks that typically require human intelligence. These systems are designed to perceive their environment, process information, learn from experience, and make decisions to achieve specific goals. 


Primary Classification 

  1. Traditional Learning (Non-Machine Learning) 
    1. Rule-Based Systems / Expert Systems: Uses predefined rules, logic, and knowledge bases; Deterministic, explainable, manually programmed knowledge in Nature 
    2. Search and Optimization Algorithms: Systematic exploration of solution spaces (e.g., Heuristic search (A*, greedy algorithms) 
    3. Fuzzy Logic Systems: Handles uncertainty and partial truth using fuzzy sets 
    4. Knowledge Representation and Reasoning: Formal representation of information and logical inference (e.g., Ontology-based systems) 
  2. Machine Learning (ML): Machine Learning is a subset of AI where systems learn patterns from data without being explicitly programmed for each specific task. 

Learning Technique wise: 

  1. Supervised Learning: Learns from labeled input-output pairs 
    1. Linear Models (Linear regression, logistic regression, support vector machines (linear kernel)) 
    2. Tree-Based Methods (Decision trees, random forests, gradient boosting (XGBoost, LightGBM)) 
    3. Ensemble Methods (Combining multiple models for better performance) 
  2. Unsupervised Learning: Finds hidden patterns in unlabeled data 
    1. Clustering: K-means, hierarchical clustering, DBSCAN 
    2. Dimensionality Reduction: Principal Component Analysis (PCA), t-SNE, UMAP 
    3. Association Rule Learning: Apriori algorithm, market basket analysis 
    4. Reinforcement Learning: Learns through interaction with environment using rewards and penalties 
      1. Classical Reinforcement Learning: Q-learning, SARSA, policy gradients 
      2. Deep Reinforcement Learning: Deep Q-Networks (DQN), Actor-Critic methods, PPO 
    5. Self-Supervised Learning 
      1. SimCLR, MoCLR 

Methodology Wise: 

  1. Deep Learning (DL): Deep Learning is a subset of Machine Learning that uses artificial neural networks with multiple layers  
    1. Multilayer Perceptrons (MLPs) 
    2. Convolutional Neural Networks (CNNs) 
    3. Recurrent Networks/Long Short-Term Memory (LSTM) / Gated Recurrent Units (GRU) 
    4. Transformer Architecture 
      1. Generative Models 
        1. Generative Adversarial Networks (GANs) 
        2. Variational Autoencoders (VAEs) 
        3. Diffusion Models 
        4. Large Language Models (LLMs) 

How Traditional AI Works: 

Traditional AI systems use one of several approaches: 

  • Rule-Based Logic: Following explicit if-then rules (e.g., "If spam score > 0.8, mark as spam") that humans have programmed 
  • Logical Reasoning: Using formal logic and knowledge bases to reach conclusions 
  • Optimization: Systematically exploring solutions to find the best outcome based on defined criteria (not learned patterns) 

Strengths: 

  • You can understand exactly why a decision was made because the rules are explicit and human-written 
  • Performs consistently on the same inputs, no surprises 
  • Works without needing large datasets; rules are programmed by experts 
  • Clear rules make them appropriate for high-stakes decisions (banking, aviation) 

Limitations: 

  • Cannot adapt or improve beyond what humans explicitly programmed 
  • Fails on situations outside the scope of programmed rules 
  • Requires domain experts to manually identify and encode all rules 

How Generative AI Works: 

Generative AI systems learn the underlying structure of data by processing billions of examples. When you provide a prompt, the system predicts what should come next based on patterns it learned during training; whether that's the next word in a sentence, the next pixel in an image, or the next note in a musical composition. 


Strengths: 

  • Can create novel, creative, and contextually relevant outputs 
  • Handles ambiguous or open-ended prompts 
  • Works well with large, diverse datasets 
  • Increasingly accessible through user-friendly interfaces 

Limitations: 

  • Can "hallucinate"; generating confident-sounding but false or misleading information 
  • May reflect biases present in training data 
  • Computationally expensive to train and run 
  • Outputs require verification and cannot always be trusted without fact-checking 
  • Raises important questions about authenticity, copyright, and originality