module

Modules 1 and 2 provided historical context on the evolution of AI and its applications in various sectors.
Module 3 delves into how AI systems work, specifically how raw data is transformed into models that perform tasks like diagnosing diseases, translating languages, recommending products, or detecting fraud.
This module is not focused on programming; instead, it aims at systems literacy, equipping professionals with a conceptual understanding of AI architecture.
Understanding AI architecture enhances professionals' abilities to:
  - Commission AI projects
  - Evaluate vendor claims
  - Identify potential failure modes
  - Ask relevant questions in technical meetings
Ultimately, participants will gain a form of technical literacy essential for engaging effectively with AI systems in their careers.

The core principle in AI: AI systems learn from data, not from explicit rules or human common sense.
Data types are categorized as:
  - Structured Data: Organized in defined formats (e.g., databases, spreadsheets). Examples include:
    - Financial transaction records
    - Patient demographic information
    - Agricultural weather data
    - Government census data
  - Unstructured Data: No predefined format (e.g., text, images, audio, video).
The majority of data generated today is unstructured, which fuels the effectiveness of deep learning systems.
Key Insight: AI project failures commonly arise from data issues such as incomplete data, inconsistent labeling, unrepresentative sampling, and poor data governance.
- Example: A model trained on biased data yields biased outputs, leading to severe consequences.

Machine Learning: Techniques allowing AI systems to learn from data instead of human-programmed rules.
Three Paradigms of Machine Learning:
  - Supervised Learning:
    - Involves training on input-output pairs (data with both input and correct output provided).
    - Adjusts internal parameters to minimize prediction error.
    - Common tasks:
      - Classification: Predicting categories
      - Regression: Predicting continuous values
  - Unsupervised Learning:
    - Training on unlabeled data, where the algorithm finds structure independently.
    - Common tasks:
      - Clustering: Grouping similar examples
      - Dimensionality Reduction: Compact representations
      - Anomaly Detection: Identifying outliers
  - Reinforcement Learning:
    - Trains agents through reward-based systems.
    - Used for game-playing AI, robotic control, and training large language models (RLHF).
Key concepts in machine learning training:
  - Training Dataset: The data used for training the model.
  - Validation Dataset: Monitors performance during training.
  - Test Dataset: Provides unbiased estimates after training completion.
Importance of distinct datasets: Evaluating on training data can yield misleadingly optimistic performance results.

Neural Networks: Fundamental structure behind modern AI.
Deep Learning: Involves using neural networks with multiple layers for advanced learning capabilities.
Components of an Artificial Neural Network:
- Neurons/Nodes: Computational units taking inputs and producing outputs after weight adjustments.
- Weight Adjustment: Achieved through backpropagation to minimize error contributions.
Hierarchical Feature Learning: Depth in neural networks allows complex feature representation.
- Example: Image networks learn simple to complex shapes across layers.
Specialized architectures for different tasks:
  - Convolutional Neural Networks (CNNs): Designed for grid-structured data handling (e.g., images).
  - Recurrent Neural Networks (RNNs): For sequential data processing (e.g., text, audio).
  - Transformers: Introduced self-attention mechanisms, highly effective for natural language processing.

Building an AI system requires more than just algorithms and data; it requires a robust ecosystem of tools and platforms.
Key Programming Language: Python, has dominated AI development for its ease of use and vast library ecosystem.
Leading Deep Learning Frameworks:
- TensorFlow: Developed by Google for scalable machine learning.
- PyTorch: Developed by Meta AI for high-performance, imperative deep learning.
Hugging Face: Platform for hosting pre-trained models and datasets, simplifying model fine-tuning and deployment.
Computational Resources: AI model training is resource-intensive (e.g., training GPT-4 cost tens of millions).
Cloud Platforms:
- Essential for providing access to computing resources (e.g., GPU, TPU).
- Critical for organizations in regions like Africa, facilitating AI capability building effectively and affordably.

Building an AI system is a multi-stage process:
  - Problem Framing: Defining the issue and assessing feasibility for AI solutions.
  - Data Collection and Preparation: May consume 60-80% of project time and budget.
  - Model Training: Selecting models, setting hyperparameters, monitoring for overfitting.
  - Evaluation: Testing on held-out data to assess model performance.
  - Deployment: Making the model available for use through various platforms.
  - Ongoing Monitoring: Tracking model performance and addressing model drift from changes in data distribution.
Importance of human judgment throughout the AI lifecycle:
- Human involvement is necessary in problem framing, data curation, model evaluation, and decision-making about outputs.
- The risk of inadequate oversight can lead to adverse outcomes in AI decision-making.

Understanding AI architecture enables professionals to engage more effectively with AI solutions and vendors.
Questions to consider when assessing AI proposals include:
  - Sources of training data
  - Labeling processes and representation of deployment context
  - Governance needs at different lifecycle stages
Importance for African Professionals:
- As AI projects grow in Africa, professionals equipped with systems literacy can critically evaluate and negotiate AI systems effectively.

Module 3 introduced essential AI architecture concepts necessary for understanding AI investments and their implications.
The five following sub-modules will explore these foundational areas in greater depth.