Comprehensive Study Notes on Artificial Intelligence Fundamentals and Ethics

Fundamental Concepts of Artificial Intelligence

• Definition of Artificial Intelligence (AI): A machine is considered to have artificial intelligence when it possesses the ability to mimic human traits, specifically making decisions, predicting future outcomes, and learning and improving autonomously.

• Core Capabilities: An AI system is characterized by its capacity to accomplish tasks independently through a cycle of collecting data, understanding it, analyzing it, learning from it, and improving its own performance.

• Broad Scope: AI is simultaneously defined as a form of intelligence, a type of technology, and an academic field of study. It involves the theory and development of computer systems—including both hardware machines and software—capable of performing tasks that traditionally necessitate human intelligence.

• Objective: The central goal of AI is to construct machines and algorithms capable of computational tasks that would otherwise require human-like brain functions. It is projected to impact every professional field in the future.

The Three Major Domains of AI

Artificial Intelligence is categorized into three primary domains based on the type of data and human-like function they replicate:

• Data for AI (Statistical Data): This domain utilizes statistical techniques to analyze, interpret, and derive insights from numerical or tabular data. It focuses on identifying patterns within large datasets to make predictions.

• Natural Language Processing (NLP): This domain focuses on textual and auditory data. It enables machines to comprehend, generate, and manipulate human language. It involves recognizing speech patterns, inferring meaning, and providing useful responses.

• Computer Vision (CV): This domain works with visual data, such as images and videos. It enables machines to interpret and understand visual information from the world, identifying objects, faces, and patterns as a human would.

Practical AI Applications and Case Studies

• Face Lock in Smartphones: Uses Computer Vision. The front camera captures and saves facial features during initialization. When the user attempts to unlock the phone, the system detects current features and compares them to the saved data; if they match, the device unlocks.

• Smart Assistants: Examples include Apple’s Siri and Amazon’s Alexa. These use NLP to recognize speech patterns, infer the user's intent, and provide an appropriate response.

• Fraud and Risk Detection: Financial companies use statistical data for customer profiling. By analyzing past expenditures and other variables, AI identifies the probability of risk or default, helping to prevent bad debts and losses.

• Medical Imaging: Computer-supported applications assist physicians by converting $2$D scan images into interactive $3$D models. This allows for detailed interpretation of a patient’s health condition and aids in diagnostic accuracy.

• Pest Management (CottonAce): An AI-enabled mobile application used in India to protect cotton crops from the Pink Bollworm. Farmers take a picture of captured pests; the app detects the insect, level of infestation, and recommends specific pesticide amounts and spray timings. Small farms using this app saw profit margin increases of up to $26.5\%$ and a drop in pesticide costs of up to $38\%$.

• Preventable Blindness (Diabetic Retinopathy): Google partnered with Aravind Eye Hospital in India to develop an AI screening solution. Automated analysis of digital retinal images achieved a detection accuracy of $98.6\%$, helping identify conditions that cause blurred vision and blindness in diabetic patients, even in rural areas without specialist doctors.

The AI Project Cycle

The AI project cycle is a modular, six-stage process used to break down complex problems into manageable steps:

1. Problem Scoping: Identifying a problem and establishing a vision to solve it.

2. Data Acquisition: Collecting relevant, authentic information from reliable sources (like government portals $data.gov.in$) to train the model.

3. Data Exploration: Visualizing data to understand trends, patterns, and relationships through various graphical representations.

4. Modelling: Selecting and training an AI algorithm (Rule-Based or Learning-Based) to perform specific tasks.

5. Evaluation: Testing the model's reliability using testing data and comparing predictions with reality.

6. Deployment: Implementing the solution in a real-world scenario, such as a mobile app or website.

Stage 1: Problem Scoping and the 4Ws Canvas

To effectively define the goal of an AI project, practitioners use the 4Ws Problem Canvas:

• Who?: Identifies the stakeholders affected by the problem and those who will benefit from the solution.

• What?: Determines the nature of the problem and identifies evidence (reports, articles) proving its existence.

• Where?: Focuses on the context, situation, or physical location where the problem occurs.

• Why?: Analyzes the benefits the solution will provide to stakeholders and society.

• Problem Statement Template: A summary tool that combines these elements: "Our [stakeholders] have a problem that [issue/need] when/while [context]. An ideal solution would [benefit]."

Stage 2: Data Acquisition and System Mapping

• Data Types:

  • Training Data: The initial dataset used to train the machine to learn patterns.

  • Testing Data: A separate dataset used to evaluate the efficiency of the trained model.

• Data Features: The specific types of data collected (e.g., for salary prediction: amount, increment period, bonus).

• System Maps: A visual tool to represent complex issues where multiple elements are interconnected.

  • Relationship Indicators:

    • A + sign indicates a direct relationship (as $X$ increases, $Y$ increases).

    • A - sign indicates an inverse relationship (as $X$ increases, $Y$ decreases).

  • Feedback Loops: Chains of cause and effect within the system.

Stage 3: Data Exploration and Visualization

Visualization is essential to quickly comprehend trends before model training. Key graphical representation types include:

• Bar Chart: For comparing quantities across different categories.

• Scatter Plot: For showing the relationship between two variables.

• Line Chart: Used for tracking changes over periods of time.

• Tree Diagram: For representing hierarchies or decision processes.

Stage 4: AI Modelling Approaches

AI serves as an umbrella term for Machine Learning ($ML$) and Deep Learning ($DL$).

• Artificial Intelligence (AI): Any technique enabling computers to mimic human intelligence.

• Machine Learning (ML): A subset of AI that allows machines to improve at tasks through experience and learning from previous mistakes.

• Deep Learning (DL): A subset of ML that uses vast amounts of data and complex neural networks to train the software to perform tasks independently.

• Classification of Approaches:

  • Rule-Based: Rules are explicitly defined by a developer. The machine follows these instructions statically. Learning does not change unless the rules are manually updated.

  • Learning-Based: The machine is fed data and desired outputs; it then designs its own algorithms. This approach is adaptive to changes and handled exceptions better (e.g., distinguishing an apple from a banana based on varying features).

Stage 5: Evaluation and Testing

• Evaluation Purpose: To understand the reliability and efficiency of an AI model using parameters like accuracy and the ROC metric.

• Overfitting: A risk where a model remembers the training data too perfectly and fails to generalize to new, unseen data.

• Performance Cases (e.g., Forest Fire scenario):

  • True Positive: The model predicts a fire, and there is a fire (Correct).

  • False Positive: The model predicts a fire, but there is no fire (Incorrect/Alarm error).

  • False Negative: The model predicts no fire, but there is a fire (Incorrect/Dangerous error).

Unit 1.3: Ethics and Morality in AI

• Difference Between Morals and Ethics:

  • Morals: Beliefs dictated by society or religion; they are not fixed and vary between cultures (e.g., "always speak the truth").

  • Ethics: Personal guiding principles or choices values that determine what is good or bad in specific situations (e.g., "Is it good to speak the truth in all situations?").

• Major Ethical Principles for AI:

  1. Human Rights: Ensuring AI does not take away freedom, discriminate, or unfairly deprive people of jobs.

  2. Bias: Avoiding partiality in training data. If data is biased, the AI results will be biased (e.g., a recruitment tool penalizing resumes containing the word "women").

  3. Privacy: Ensuring personal data is not extracted or used without consent.

  4. Inclusion: Ensuring AI does not leave out specific populations (e.g., ensuring both rich and poor benefit equally).

• The Moral Machine: A platform designed to gather human perspectives on moral decisions made by AI, such as how a self-driving car should react when faced with unavoidable accidents (choosing between protecting passengers or pedestrians).