Introduction to AI Flashcards
Artificial intelligence
General Types of AI
Artificial intelligence (AI) refers to computer systems capable of performing complex tasks that historically only a human could do, such as reasoning, making decisions, or solving problems.
General Types of AI
By Capability
* Narrow AI task-specific
* General AI human-level intelligence
* Superintelligent AI beyond human intelligence
By Functionality
* Reactive machines no memory
* Limited memory learns from past experiences
* Theory of mind understands emotions
* Self-aware AI hypothetical consciousness
By Approach:
- Rule-based AI
- Machine learning-based AI
- Deep learning-based AI
- Natural language processing AI
General Types of AI – Capability
General Types of AI – Capability
Narrow AI: task-specific
- Definition: AI systems designed to perform a specific task or set of tasks.
- Examples: Virtual assistants (Siri, Alexa), Recommendation algorithms (Netflix, Amazon), Facial recognition software, and
Chatbots - Key Points: Narrow AI cannot perform tasks outside of its trained domain and lacks general intelligence.
General AI human-level intelligence
- Definition: An AI system with generalized intelligence, meaning it can perform any intellectual task that a human can do.
Does not exist yet. - Examples: Would be comparable to human cognition across various domains
- Key Points: General AI would be able to learn, reason, and solve problems across a wide range of fields without being pre-programmed
Superintelligent AI: beyond human intelligence
- Definition: A theoretical form of AI that surpasses human intelligence in all respects, including creativity, problem- solving, and decision-making.
- Examples: Does not exist yet.
Would be beyond human cognition across domains - Key Points: would potentially redefine technology, but it raises ethical concerns about control and safety.
General Types of AI – by Functionality
Reactive machines: no memory
- Definition: Simplest form that reacts to specific inputs without ability to form memories or use past experiences to influence future decisions.
- Examples: IBM’s Deep Blue, the chess-playing AI.
- Key Points: Reactive machines can handle specific tasks but do not learn or adapt over time.
Limited memory learns from past experiences
* Definition: Use past experiences and historical data to make better decisions, but information is not permanently stored.
- Examples: Self-driving cars that use data about traffic patterns or obstacles.
- Key Points :Limited-memory AI systems learn from historical data but are still specialized and task-specific.
Theory of mind understands emotions
- Definition: More advanced type that, in theory, could understand emotions, intentions, and human mental states
Not yet fully developed. - Examples: AI that could engage in more human-like interactions by understanding user emotions is in progress.
- Key Points:Would be better at social interactions and complex decision-making but remains largely theoretical.
Self-aware AI hypothetical consciousness
- Definition: Theoretical system that has consciousness, self-awareness, and emotions like humans.
- Examples: Does not exist yet.
- Key Points: Self-aware AI would raise philosophical, ethical, and existential questions about its role in society.
General Types of AI – by Approach
Rule-based AI
- Definition: Follows predefined rules and logic to make decisions. Requires explicit instructions for decision-making.
- Examples: Early expert systems in medicine or finance that use if-then rules.
- Key Points: Rule-based AI is highly interpretable but limited in its ability to adapt or learn from new data.
Machine learning-based AI
- Definition: Learns from data and improves over time without being explicitly programmed.
Identifies patterns and makes predictions or decisions. - Examples: Spam filters, recommendation systems, and AI in video games
- Key Points: Allows AI to adapt to new situations, making it more flexible than rule-based systems.
Deep learning-based AI
- Definition: A subset of machine learning that uses artificial neural networks to learn from vast amounts of data. Particularly effective for image recognition, speech processing, and natural language understanding.
- Examples: Google’s AlphaGo, deep neural networks for image classification.
- Key Points: require large datasets and computing power but are capable of handling more complex tasks.
Natural language processing AI
- Definition: Specialized AI focused on enabling machines to understand, interpret, and respond to human language.
- Examples: Chatbots, translation tools, and virtual assistants.
- Key Points: NLP involves tasks like sentiment analysis, language translation, and speech recognition.
General Types of AI – Overviews
General Types of AI – Overviews
S25 - 28 Folie 03
General Types of AI – Reinforcement Learning (I)
Advantages
- Solves higher-order and complex problems. Will be very accurate.
- Reason: very similar to the human learning technique.
- Rigorous training process, takes time but helps to correct any errors.
- Learning ability – can be termed as deep reinforcement learning.
- Model learns constantly, a earlier mistake unlikely in the future.
- Various problem-solving models are possible.
- Even when no training data, it will learn through experience.
- For various problems, which might seem complex to us, it provides the perfect models to tackle them.
Disadvantages
- For solving simpler problems not correct.
- Requires unnecessary processing power and space
- We need lots of data. Reinforcement Learning models require a lot of training data to develop accurate results.
- Maintenance costs are very high.
- Like for driverless vehicles, robots, we would require a lot of
maintenance for both hardware and software. - Excessive training can lead to overloading of the states of the model.
System Architecture of AI
S32-39 Folie 03
History and Evolution of AI
1950s - Early Foundations:
* 1950: Turing Test
Alan Turing proposes a test to determine if a machine can exhibit intelligent behavior indistinguishable from a human.
* 1956: Dartmouth Conference
Considered the birth of AI as a field of study. Researchers, including John McCarthy, coined the term “Artificial Intelligence.”
* 1957: Perceptron Model
Frank Rosenblatt develops the first neural network, laying the groundwork for machine learning.
1960s – AI in Reasoning and Problem-Solving
* 1961: UNIMATE
The first industrial robot, introduced in General Motors, showed early automation in manufacturing.
* 1966: ELIZA
Joseph Weizenbaum develops ELIZA, an early natural language processing program mimicking human conversation (early chatbot)
* 1969: Shakey the Robot
The first general-purpose robot, capable of perceiving its environment and performing simple tasks autonomously.
1970s – Expert Systems and Early AI Applications:
* 1972: PROLOG
Logic programming language for AI development, widely used in AI research, especially in natural language processing and problem- solving.
* 1979: Stanford Cart
One of the earliest self-driving vehicles, it successfully navigated obstacles autonomously.
1980s – AI in Commercial Use:
* 1980s: Expert Systems
AI systems like MYCIN (medical diagnosis) and XCON (configuring computer systems) dominated industries, using rule-based logic to solve complex problems.
* 1986: Backpropagation Algorithm
Geoffrey Hinton popularizes backpropagation, revolutionizing neural networks and laying the foundation for modern deep learning.
1990s – A I Achieves High-Profile Wins: * 1997: IBM’s Deep Blue
Deep Blue defeats world chess champion Garry Kasparov.
2000s – AI Expands into New Domains:
* 2000: Kismet
A robot developed at MIT with emotional AI capabilities, Kismet could recognize and respond to human emotions.
* 2002: Roomba
The introduction of the Roomba robot vacuum showed how AI could be incorporated into consumer products.
2010s – Deep Learning Revolution:
* 2011: IBM Watson
Watson wins Jeopardy!, demonstrating AI’s ability to process and understand natural language at a high level.
* 2012: AlexNet
This deep learning model wins the ImageNet competition, marking a turning point in image recognition and computer vision.
* 2016: Google DeepMind’s AlphaGo
AlphaGo defeats the Go world champion, showcasing AI’s potential in complex, intuitive games.
2020s – AI and Generalization:
* 2018: GPT-2
OpenAI introduces a highly advanced natural language processing model, capable of generating human-like text.
* 2020: GPT-3
GPT-3, one of the most powerful natural language models to date, capable of understanding & generating human-like text with minimal input.
* 2023: ChatGPT (based on GPT-4)
A conversational AI model capable of assisting with a wide range of complex tasks, from writing to problem-solving.
Key Approaches / Types of Algorithms in AI
Symbolic AI (1)
- Symbolic AI (Rule-Based AI): Explicitly encodes knowledge in form of rules.
- Machine Learning: Data-driven models learning patterns and making predictions.
- Deep Learning: Advanced neural networks for tasks like image and speech recognition.
- Evolutionary Algorithms: Natural selection principles to evolve solutions.
- Bayesian Inference: Probabilistic models that handle uncertainty.
- Reinforcement Learning: Agents learn through trial and error by interacting with their environment.
- Cognitive Computing: Mimics human thought processes & decision-making.
- Hybrid Approaches: Combine multiple AI techniques to create more powerful systems.
Symbolic AI (1)
- Definition: Based on explicitly encoding human knowledge: Rules & logic. “Good Old-Fashioned AI” (GOFAI).
- Key Concepts:
- Knowledge Representation: Information is encoded in symbolic form,
such as facts, rules, and logical structures. - Logic and Inference: AI systems use logical reasoning to draw conclusions from the encoded knowledge.
Examples
- Expert Systems: AI programs that apply a set of rules to data in order to make decisions, such as MYCIN (medical diagnosis) and DENDRAL (chemical analysis).
- Decision Trees: A hierarchical structure where decisions are made based on logical rules.
- Strengths: Easy to interpret and explain.
- Weaknesses: Hard to scale to more complex, unstructured data; inflexible to new situations.
Machine Learning (1)
- Definition: Machine learning (ML) enables computers to learn from data and make decisions or predictions based on patterns. It
is data-driven rather than relying on pre-programmed rules.
Key Concepts:
- Supervised Learning: AI learns from labeled examples, where input-output pairs are known.
- Unsupervised Learning: AI finds hidden patterns in data without any labeled outputs.
- Reinforcement Learning: AI learns by interacting with its environment & receiving feedback (rewards or penalties).
Example Systems:
- Spam Filters: Supervised learning algorithms used to classify emails as spam or not.
- Customer Segmentation: Unsupervised learning used for clustering customers based on purchasing behavior.
- Game-Playing Agents: AlphaGo, which learn to play complex games through trial and error.
Strengths: Highly adaptable, handles large amounts of data, can generalize to new situations.
Weaknesses: Requires large datasets, may not always provide interpretable results (black box problem).
Deep Learning
Definition: Subset of machine learning that uses neural networks with multiple layers to model complex, high-dimensional
patterns in data.
Key Concepts:
- Neural Networks (NN): Layers of interconnected “neurons” that process input data and extract features.
DL has many layers (deep neural networks). - Convolutional NN (CNNs): For image and video analysis, particularly in image recognition and classification.
- Recurrent NN (RNNs): Specialized for sequential data, such as time-series data or natural language.
Example Systems:
- Image Classification: CNNs used in applications like facial
recognition or autonomous driving. - Natural Language Processing (NLP): RNNs and transformer-based models like GPT-3 for text generation and understanding.
Strengths: Excels at complex tasks like image and speech recognition, can process unstructured data.
Weaknesses: Requires significant computational power and large datasets, and models are often difficult to interpret.
Evolutionary Algorithms
Definition: Inspired by natural selection, these algorithms evolve solutions to optimization problems over time by generating and testing variations of possible solutions.
Key Concepts:
* Genetic Algorithms: Use principles of mutation, crossover (recombination), and selection to evolve solutions.
- Fitness Function: Measures how well a solution solves the problem.
Example Systems:
- Neural Network Optimization: Evolutionary algorithms are used to evolve
neural network architectures or weights. - Optimization Problems: Applied in fields such as engineering design and finance for optimizing solutions.
Strengths: Useful for where the search space is vast and complex.
Weaknesses: Often computationally expensive and slow, especially for large-scale problems.
Bayesian Inference
Definition:
Uses probability theory to make inferences and predictions based on uncertainty. It relies on Bayes’ Theorem to update the probability of a hypothesis as more data becomes available.
Key Concepts:
- Bayesian Networks: Graphical models representing probabilistic relationships between variables.
- Hidden Markov Models (HMMs): Used to model systems where the underlying state is hidden, but observable outcomes are available.
Example Systems:
- Spam Detection: Bayesian inference is commonly used to classify emails based on the probability of containing spam.
- Speech Recognition: HMMs are used in systems where the sequential nature of spoken language needs to be modeled.
Strengths: Effective for handling uncertainty and making probabilistic predictions.
Weaknesses: Can become complex and computationally expensive for large-scale problems with many variables.
Reinforcement Learning
Definition: An agent learns to make decisions by interacting with an environment & receiving feedback (rewards or penalties). Goal is to maximize cumulative rewards over time.
Key Concepts:
* Q-Learning: A value-based approach where the agent learns the value of different actions in specific states.
* Policy Gradient Methods: The agent directly learns a policy for action selection without using a value function.
Example Systems:
* Autonomous Robots: Robots learn to navigate and interact with their environments through trial and error.
- Game AI: AlphaGo, a reinforcement learning agent that learned to play Go at a superhuman level.
Strengths: Ideal for problems where sequential decision-making is required, can learn from its own experience.
Weaknesses: Requires a lot of trial and error, making training slow and resource-intensive.
Cognitive Computing
Definition:
This approach aims to mimic human thought processes and decision-making by simulating how the human brain
works. Cognitive computing systems are designed to understand, reason, and learn in a more human-like way.
Key Concepts:
* Natural Language Processing (NLP): Systems that can understand and generate
human language
- Contextual Understanding: Cognitive systems take context into account when making decisions or solving problems.
Example Systems:
* IBM Watson: A cognitive computing system that understands and processes vast amounts of information, such as medical research and patient data.
Strengths: Capable of tackling complex decision-making tasks, particularly in domains like healthcare and finance.
Weaknesses: Requires significant computing resources, and fully mimicking human cognition is still an ongoing challenge.
Hybrid Approaches
Definition:
These approaches combine multiple AI techniques to build more robust, efficient systems.
E.g., ML can be combined with rule-based AI, or NNs can be used alongside probabilistic reasoning models.
Key Concepts:
* Multi-Agent Systems:Systems that involve multiple AI agents working together or in competition.
- Deep Reinforcement Learning: A combination of deep learning and reinforcement learning to create systems capable of handling complex environments (e.g., autonomous vehicles).
Example Systems:
* Robotics: Combining machine learning with rule-based logic to control autonomous robots in dynamic environments.
Strengths: Allows systems to leverage the best aspects of different approaches, improving performance and flexibility.
Weaknesses: Complex to design and implement due to the integration of multiple techniques.
Example Applications of AI
What you can do with AI
* China: Ping An – Car Insurance Claims
* China: AI in Farming
* China: Ping An Good Doctor
AIHH-Startups:
* Check for Pet
* Exazyme
* Yoona
* Streamboost
AIHH-Startups – Health Care:
* Nosc AI – Holistic Patient-Centered Practice Organization
* Fuse AI – Prostate Cancer Detection
* PDV – AI-powered Heart Check
* Casuu - Nursing Education
