Domain 1: foundations of AI Flashcards
AI
Machines performing tasks normally thought to require human intelligence
Turing Test
a machine is intelligent if it can fool humans into thinking is human
Socio-technical systems
humans shape AI while AI shapes humans
Machine Learning (ML)
sub-field of AI that uses algorithms to learn from input data, makes decisions/inferences/predictions, and performs tasks on new data without being explicitly programmed to do so
Supervised learning overview
- Requires large labeled datasets (5k+ for images)
- Human labeled (knowingly or unknowingly)
Large-scale AI Open Network (LAION)
Has lots of large labeled datasets to use for training
Supervised learning techniques (4)
Classification models (cat or dog)
Regression models (vehicle fuel efficiency)
Support Vector Machine (SVM): classification
Support Vector Regression (SVR): regression
Unsupervised learning overview
extract features from data without labels, making the algorithm less predictable
Unsupervised sub-cateogries
Clustering and association rule learning which doesn’t require the overhead of labeling data
Supervised vs unsupervised
- Neither is better
- For financials:
Supervised: find known types of fraud
Unsupervised: find new patterns of behavior
Semi-supervised learning
Combines the benefits of both supervised and unsupervised learning to improve reliability and reduce cost of training.
Example: Large Language Models (LLMs)
Reinforcement Learning
simulated motivation through rewards and punishment with no pre-labeled data
Uses a learning feedback loop where data is gathered through interaction with the environment
Learning feedback loop
Action > feedback > repeat/adjust action
- Action > reward > continue/refine
- Action > punishment > change action
rewards need to be proportionate to the action
Reinforcement Learning challenges
Works well for games
Real life is not as heavily bound as games
Time cannot be sped up
Failure may have serious consequences (self-driving cars)
Discriminative model
classification technique to determine how data is grouped and create decision boundaries that cause the grouping
Example: label animals by type
Generative model
generates requested data based on underlying known characteristics
Example: GenAI creating a picture of a dog
Types of foundation models
Large Language Models (LLMs)
Vison models: interpret visual data
Scientific models: protein folding
Audio models: generate human speech, music
Value of Foundational models
- they save time and resources
- generalized, adaptable, and scalable models
- enable transfer learning, fine tuning
Categories of AI (3)
Artificial Narrow Intelligence (ANI)
Artificial General Intelligence (AGI)
Artificial Super Intelligence (ASI)
Expert Systems components
- Knowledge base: organized facts for specific domains
- inference engine: rules-based system to locate facts for a prompt
- User interface: user inputs lead to system outputs
Fuzzy logic system (4 steps)
- Fuzzification: input converted to fuzzy data
- Rule eval: matches input to rule
- Aggregation: rule outputs combined
- Defuzzification: fuzzy outputs converted back to specific values
Example: automate vehicle breaking gets stronger as the car gets “closer”
Common AI models (4 main groups, 3 sub groups)
Linear and statistical models
Decision trees
Robotics
Machine Learning > Neural networks >
- Computer Vision
- Speech recognition
- Language models
ML == hidden layers
Types of neural networks (3)
Computer vision
Speech recognition
Natural Language Processing (NLP)
AI use cases (Mnemonic)
Rabid = Recommendation
Raccoons = Recognition
Dance = Detection
Fancifully with = Forecasting
Giant = Goal-driven optimization
Indigo = Interaction support
Penguins = Personalization
