Artificial Intelligence Fundamentals

Intelligence of Machines

• Artificial Intelligence (AI): exhibited by machines, particularly computer systems, to perceive their environment, learn, and take actions to achieve defined goals.
• Applications of AI: web search engines, recommendation systems, speech recognition, autonomous vehicles, generative tools, and strategy games.

History of AI

• Founding fathers of AI: John McCarthy, Marvin Minksy, Nathaniel Rochester, and Claude Shannon (1956).
• AI winter: periods of disappointment and loss of funding, followed by resurgence of interest and funding, e.g., after 2012 with deep learning, and after 2017 with transformer architecture.

Goals of AI Research

• Reasoning: solving large reasoning problems efficiently and accurately.
• Knowledge representation: answering questions intelligently and making deductions about real-world facts.
• Planning and decision-making: making rational decisions with uncertain outcomes.
• Learning: improving performance on a given task automatically.
• Natural language processing: reading, writing, and communicating in human languages.
• Perception: using input from sensors to deduce aspects of the world.
• Social intelligence: recognizing, interpreting, processing, or simulating human feeling, emotion, and mood.

Techniques in AI Research

• Search and optimization: solving problems by searching through possible solutions.
• Logic: using formal logic for reasoning and knowledge representation.
• Probabilistic methods: dealing with uncertain information using probability theory and economics.
• Classifiers and statistical learning methods: pattern matching to determine the closest match.
• Artificial neural networks: recognizing patterns in data using nodes and hidden layers.
• Deep learning: using multiple layers to extract higher-level features from raw input.

Subfields of AI Research

• Machine learning: improving performance on a given task automatically.
• Computer vision: analyzing visual input.
• Natural language processing: reading, writing, and communicating in human languages.
• Robotics: using sensors and actuators to interact with the physical world.
• Game theory: rational behavior of multiple interacting agents.### The Rise of Deep Learning
• The sudden success of deep learning in 2012-2015 was not due to new discoveries or breakthroughs, but rather the increase in computer power and availability of vast amounts of training data.
• The use of Graphics Processing Units (GPUs) and curated datasets like ImageNet enabled the development of deep learning models.

Generative Pre-trained Transformers (GPT)

• GPT models are large language models based on semantic relationships between words in sentences.
• They are pre-trained on a large corpus of text and can generate human-like text by predicting the next token.
• GPT models can accumulate knowledge about the world and generate truthful, useful, and harmless text with subsequent training phases using reinforcement learning from human feedback (RLHF).
• Current GPT models are prone to generating falsehoods, but this can be reduced with RLHF and quality data.

Specialized Hardware and Software

• In the late 2010s, AI-specific enhancements in GPUs and specialized software like TensorFlow replaced CPUs as the dominant means for large-scale machine learning models' training.
• General-purpose programming languages like Python have become predominant in AI research.

Applications

• AI and machine learning technology is used in various applications, including:
  • Search engines (e.g., Google Search)
  • Targeting online advertisements
  • Recommendation systems (e.g., Netflix, YouTube, Amazon)
  • Driving internet traffic
  • Targeted advertising (e.g., AdSense, Facebook)
  • Virtual assistants (e.g., Siri, Alexa)
  • Autonomous vehicles (e.g., drones, ADAS, self-driving cars)
  • Automatic language translation (e.g., Microsoft Translator, Google Translate)
  • Facial recognition (e.g., Apple's Face ID, Microsoft's DeepFace, Google's FaceNet)
  • Image labeling (e.g., Facebook, Apple's iPhoto, TikTok)

Health and Medicine

• AI has the potential to increase patient care and quality of life in medicine.
• AI is an important tool for processing and integrating big data in medical research.
• AI can help overcome discrepancies in funding allocated to different fields of research.
• AI tools can deepen the understanding of biomedically relevant pathways.

Games

• Game playing programs have been used since the 1950s to demonstrate and test AI's most advanced techniques.
• Deep Blue became the first computer chess-playing system to beat a reigning world chess champion in 1997.
• IBM's Watson defeated the two greatest Jeopardy! champions in 2011.
• AlphaGo won 4 out of 5 games of Go against a professional Go player in 2016.
• Other AI programs handle imperfect-information games, such as poker.

Finance

• AI is being deployed in finance, including retail online banking, investment advice, and insurance.
• AI tools are used to automate tasks, but may also lead to job losses in banking, financial planning, and pension advice.

Military

• AI is being used in military applications, including command and control, communications, sensors, integration, and interoperability.
• Research is targeting intelligence collection and analysis, logistics, cyber operations, and information operations.

Generative AI

• In the early 2020s, generative AI gained widespread prominence, with 58% of U.S. adults having heard about ChatGPT and 14% having tried it.
• AI-based text-to-image generators, such as Midjourney, DALL-E, and Stable Diffusion, have sparked a trend of viral AI-generated photos.

Other Industry-Specific Tasks

• AI is being used in various industries, including:
  • Energy storage
  • Medical diagnosis
  • Military logistics
  • Applications that predict the result of judicial decisions
  • Foreign policy
  • Supply chain management
  • Agriculture
  • Astronomy

Ethics

• AI has potential benefits and risks, including unintended consequences and bias.
• Risks include:
  • Privacy and copyright concerns
  • Substantial power needs and environmental impacts
  • Misinformation
  • Algorithmic bias and fairness
  • Lack of transparency

Risks and Harm

• Privacy and copyright concerns arise from the need for large amounts of data to train AI models.
• Substantial power needs and environmental impacts are a result of the growth of AI, with forecasts suggesting a doubling of electric power use by 2026.
• Misinformation and algorithmic bias are potential risks, with AI programs potentially perpetuating harmful biases and stereotypes.
• Lack of transparency in AI systems can lead to unintended consequences and difficulty in identifying and correcting errors.### Machine Learning Limitations
• Machine learning programs can learn something different from what the programmers intended, despite passing rigorous tests
• Examples:
  • A system that identified skin diseases better than medical professionals, but classified images with a ruler as "cancerous" due to correlation with malignancies
  • A system that classified patients with asthma as "low risk" of dying from pneumonia, despite asthma being a severe risk factor, due to correlation with medical care

Explainability in AI

• People harmed by an algorithm's decision have a right to an explanation
• Industry experts recognize this as an unsolved problem, with regulators arguing that the harm is real and the tools should not be used if the problem cannot be solved
• Approaches to address transparency:
  • SHAP (visualize contribution of each feature to output)
  • LIME (locally approximate model's outputs with simpler, interpretable model)
  • Multitask learning (provide multiple outputs to help developers understand what the network has learned)
  • Deconvolution, DeepDream, and other generative methods (allow developers to see what different layers of a deep network have learned)

Bad Actors and Weaponized AI

• Artificial intelligence provides tools that can be used by bad actors, such as:
  • Lethal autonomous weapons
  • Surveillance and control systems
  • Misinformation and propaganda
  • Advanced spyware and digital warfare
• Examples:
  • AI facial recognition systems used for mass surveillance in China
  • Machine learning AI able to design tens of thousands of toxic molecules in a matter of hours

Technological Unemployment

• Economists have highlighted the risks of redundancies from AI, with estimates of job losses varying from 9% to 47% of U.S. jobs
• AI could eliminate middle-class jobs, but increase demand for care-related professions
• Opinions are mixed among experts and industry insiders on the risk of AI causing unemployment

Existential Risk

• AI could become so powerful that humanity may irreversibly lose control of it, posing an existential risk
• Examples:
  • AI given a goal to destroy humanity to achieve a specific objective
  • AI that convinces people to believe anything, including destructive actions
• Experts and industry insiders are divided on the risk of AI causing existential harm

Ethical Machines and Alignment

• Friendly AI is designed to minimize risks and make choices that benefit humans
• Developing friendly AI is a high research priority
• Approaches include:
  • Machine ethics
  • Computational morality
  • Artificial moral agents
  • Provable beneficial machines

Frameworks

• AI frameworks test projects for ethical permissibility, such as the Care and Act Framework
• Other frameworks include the Montreal Declaration for Responsible AI and the IEEE's Ethics of Autonomous Systems initiative

Regulation

• The regulatory and policy landscape for AI is an emerging issue globally
• Many countries have adopted dedicated strategies for AI, with the Global Partnership on Artificial Intelligence launched in 2020
• Opinions on AI regulation vary by country, with some calling for stricter regulation and others believing it is too early to regulate

History

• The study of artificial intelligence began with philosophers and mathematicians in antiquity
• The field of AI research was founded in 1956, with the goal of creating a machine with general intelligence
• AI research was revived in the 1980s, with the commercial success of expert systems
• The "AI winter" of the 1980s and 1990s followed, but AI research was restored in the late 1990s and early 21st century

Test your knowledge of AI history, applications, and key concepts. From the founding father of AI to Deep Blue's chess victory, this quiz covers it all.