Foundations of Artificial Intelligence

types of AI

1) artificial narrow intelligence (ANI) 2) artificial general intelligence (AGI) strong, deep or full AI 3) artificial superintelligence (ASI)

AI platform can

1) centralize data analysis 2) streamline development and productions workflows 3) facilitate collaboration 4) automate systems development tasks 5) monitor models and systems in production

modern drivers of AI and data science

1) cloud computing 2) mobile tech and social media 3) IOT 4) PETs and blockchain 5) computer vision, AR/VR, metaverse

3 main elements of expert system

1) knowledge base 2) inference engine 3) user interface

common AI models

1) linear and statistical 2) decision trees 3) ML 4) neural networks 5) robotics

OECD 5 dimensions developed to classify AI systems

1) ppl and planet 2) econ context 3) data and input 4) AI model 5) tasks and output

use cases and benefits of AI

1) recognition 2) detection 3) forecasting 4) personalization 5) interaction support 6) goal driven optimization 7) recommendation

3 main types of ML

1) supervised learning 2) unsupervised learning 3) reinforcement learning

primary ML models

1) supervised learning 2) unsupervised learning 3) reinforcement learning 4) semi-supervised learning

Dartmouth conference

AI born 1956 before Dartmouth, isolated disciplines, psychology, computer science, linguistics and engineering formed bedrock proposal -every aspect of learning or any other feature of intelligence can in principle be so precisely described that machine can be made to simulate it outcome/significance: term AI adopted giving birth to field of research and created new, collective sense of mission among attendees to develop machines capable of simulating human intelligence

OECD 5 dimensions to classify AI systems: econ context

AI system looked at according to econ and sectoral environment in which operates characteristics include: -sector where AI system operates -actual business function/model for AI system -necessity of AI system to operations -how deployed and impact on deployment -scale of system -tech maturity of AI system (newer system may not been tested on as much data over time; more mature systems may be more effective)

reinforcement learning

AI system rewarded for performing task well and penalized for not performing well over time learning to max rewards and develop system that works

risks in use of AI

AI systems implemented in vast, complex environments data used and AI change over time - model may need to change, upgrade, or both, to reflect what being done w/ new data in environment

artificial superintelligence (ASI)

AI systems w/ intellectual powers beyond those of humans across comprehensive range of categories and fields of endeavor capable of outperforming humans, self aware, understanding human emotions and experiences and would have ability to evoke own, thus experiencing reality like humans does not exist yet

semi-supervised learning models (ML)

addition to 3 primary types of ML models, combo of supervised and unsupervised learning processes use small amt labeled data and large amt unlabeled aimed leverage benefits of both models: improve reliability and reduce costs helpful in scenarios where challenging find/create large prelabeled datasets little more predictable than unsupervised

3 main elements of expert systems: user interface

allows end user interact w/ expert system by providing it input (problem or question) and obtaining output (resolution)

machine learning

branch of AI that leverage data and algorithms to enable systems to repeatedly learn and make decisions improve over time w/o being explicitly instructed or programmed to do so categorized based on type of training model rely on

broad artificial intelligence

combo of ANI, very specific tasks at high level efficiency more advanced in scope than ANI, capable performing broader set of tasks relies on group of AI systems, capable of working together and combining decision making capabilities (e.g., autonomous driving vehicles) lacks full, human like capabilities experts expect of AGI

linear and statistical models

common AI model model relationship btwn 2 variables e.g., how sales of product related to changes in pricing based on historical data not black box algorithm and more explainable

decision tree models

common AI model predict outcome based on flowchart of Q&A explainable and not black box disadvantage: changing training data (even in small way) can significantly impact algorithm; subject to security attacks and hacks

ML models

common AI models black box capabilities, most advanced lack of transparency and explainability neural networks (based on human brain) -contain nodes like neurons that are in layered structure and continuously improve ability to find right answer -do not need to be trained to make complex nonlinear inferences in unstructured data -commonly behind tech like FRT

neural networks

common AI models computer vision models: recognize images and videos speech recognition models: used in Alexa, transcription software, analyze speech across factors like pitch, tone, language, accent language models: NLP, allow computers understand human language using ML, deep learning and linguistics, used to process and respond to large amts of data reinforcement learning models: train models to optimize actions within given environment to achieve specific goal, guided by feedback mechanisms of rewards/penalties conducted thru trial and error, interactions/simulated experience that do not require external data, lack of explainability and transparency

robotics models

common AI models multidisciplinary field encompassing design, construction, operation and programming of robotics allows AI systems and software to interact w/ physical world w/o human intervention

understanding the AI tech stack: main areas

computer: -serverless flow -high performance -quantum computing -trusted execution environments stage and network software development

OECD 5 dimensions to classify AI system: AI model

discusses technical type how model built and used

unsupervised learning models (ML)

do not rely on labeled data sets designed ID differences, similarities and other patterns by selves w/o human supervision most cost efficient and require less effort; susceptible producing less accurate outputs and can display unpredictable behaviors 2 categories 1) clustering 2) association rule learning

fuzzy logic systems

employ fuzzy inference mechanisms to make decisions based on fuzzy rules and input data follow 4 standard steps 1) fuzzification: input data converted into fuzzy datasets 2) rule evaluation: determines degree of matching btwn rules and input data 3) aggregation: rule outputs combined 4) defuzzification: process thru which fuzzy outputs converted back into specific values

familiarity w/ common AI models and systems allow better

enable application of sound governance practices and ensure ethical and responsible uses of AI recognize different strengths and limits, potential risks and mitigations of models and systems designed for specific tasks and apps understand how each model works and how work together and whether new risks introduced effectively govern and manage AI within org products, as complex tasks may require multiple types of models to complete mission

machine perception

evolving field of potential convergence btwn robotics and AI systems trained process sensory info and mimic human senses robotic sensors provide relevant data thru cameras, microphones, pressure sensors, 3D scanners, motion detectors and thermal imaging combining sensors and AI models enable systems to sift thru data at much faster rate and order of magnitude beyond human ability eliminating noise, analyzing and categorizing info

robotic process automation (RPA)

evolving tech that uses software robots to automate repetitive and rule based tasks within business processes designed mimic human actions on digital systems (e.g., data entry and form processing) natural language processing or ML enhances RPA robot automation capabilities

artificial general intelligence (AGI) strong, deep or full AI

intended closely mimic human intelligence remains beyond reach at present - closer to achieving development thru tech advancement experts expect have strong generalization capabilities, ability to think, understand, learn and perform complex tasks and achieve goals in different contexts and environments

supervised learning

labeled data that is grouped or classified into categories via AI system used for tech recognition, detecting spam in email, etc.

artificial narrow intelligence (ANI)

limited in scope, good for automating tasks designed perform single or narrow set of related tasks at high level of proficiency operate under narrow constraints and limits commonly referred to as weak AI boost productivity and efficiency by automating repetitive tasks, enabling smarter decision making and optimization thru trend analysis

linguistic variables and fuzzy rules

linguistic variables describe concepts in natural language terms like low, medium or high and warm, hot or very hot fuzzy rules express relationships btwn variables by using if-then stmts e.g., rule may state that if temp hot then fan speed should be set to high

what is artificial intelligence?

machines performing tasks that normally require human intelligence

storage

main area of infrastructure 4 general stages of AI: 1) ingestion 2) preparation 3) training 4) output (inference) each stage has different storage requirements that must be adhered to in order avoid project failure consider: -expense for massive amt data -storage for variety storage types; each require different storage subsystems -types for structured and unstructured (easier process structured, API done to scale)

software

main area of infrastructure driven and contributed to AI challenges -democratization of AI -training AI systems -scale AI models -transformation -labeling fine tune and customize

networks

main area of infrastructure high speed networks needed to support AI models: complex AI models, deep learning, NLP, LLM deliver to training data in time to AI algorithm as well as train and inference at scale thru high speed networks high performers can compute, underlying infrastructure housed in same data centers, usually in same rack and connected via fiber connections edge computing: IOT communication or network protocols based on congestion free design, esp LLM and neural networks transmission control protocol (standard) but requires packet to be sent

compute

main area of infrastructure look @ thru central processing units (CPUs) and graphical processing units (GPUs) GPUs thrust AI mvmt forward -> specialized chips that offload a lot from CPUs which allow machines be more efficient -better performance and match to algorithmic audiences -better in matching hardware to AI model for optimal performance

trusted execution environments

main area of infrastructure: compute good from privacy perspective for AI as human influence taken out of equation

serverless flow

main area of infrastructure: compute not limited to particular server or piece of hardware running code on # of hardware devices providing 2 important services or functionalities for AI 1) loose coupling: taking data from variety of sources (no longer stuck to single source) 2) scalability: running multiple instances of that code bc it's not tied to given server which helps drive AI forward

quantum computing

main area of infrastructure: compute processing data in 3 dimensions horizontally and vertically

high performance compute

main area of infrastructure: computer create isolated clusters of compute power high speed networking, specialized chipsets

observability and monitoring challenges

main area of infrastructure: software data integrity: data trained on is not relevant or close enough to data that analysis based on and may lead to inconsistent results data drift: similar to integrity; training on specific type of data and then algorithm applied to distinct other type of data; may produce faulty predictions lead to series of future mistakes and provide inaccurate data pipelines bias and discrimination: bias common, since humans creating AI algorithms may inherit bias

role of open source AI frameworks

main area of infrastructure: software maker culture allows companies greater freedom to innovate w/ AI create own massive feedback loop that drive free spread of ideas for applying AI transformation, tuning best practices and turning ideas into viable businesses or assets for orgs open source does not rely on proprietariness of product, as long as given credit can use for whatever want

observability and monitoring

main area of infrastructure: software observability intended monitor overall health and status of org data ecosystem subcomponent of data observability is AI observability and focused on monitoring performance of AI algorithm, data going in and coming out of AI algorithm and metrics of AI system data and AI observability key to success of any AI project: provides indices and metrics for performance, an in depth analysis of AI data and models and gives capability to investigate, resolve and prevent AI model issues perform outcome validation to ensure that desired outcomes delivered, align w/ AI model and can be predicted from tuning and transform

democratization of AI

main area of infrastructure: software - challenge AI easier to use, low code software, simpler AI interfaces

tuning AI systems

main area of infrastructure: software - challenges allow customization of AI model to generate more accurate outcomes and provide highly valuable insights into data usually done thru trial and error by changing hyper parameters (very numerous in complex models), varies on model type and complexity

transformation

main area of infrastructure: software - challenges data must be transformed into AI model - increase data compatibility: some AI pipelines require transforming data for compatibility w/ data the AI model will be analyzing, optimizing data quality - low data quality poses challenge: transformation may have to be done internally to model, in processing external to model, or on post-processing of model, the output

labeling

main area of infrastructure: software - challenges enriches data used for deployment, training and tuning impactful and determines quality of AI model and results; data labels need to be of a high quality and standard challenges: - low quality labels, scaling high quality data, data labeling operations, and lack quality assurance within data labeling operations; requires one to verify and provide validation that data labels are high quality

scale AI models

main area of infrastructure: software - challenges trial and error tuning (doing it as you go) wants to move from trial and error to more educated tuning

fuzzy logic

method of reasoning intended to mimic or resemble human decision making conventional logic used in computing (crisp logic) generally take form of precise inputs and outputs, often binary in nature like true/false, yes/no enable range of possible inputs to achieve output; allows for situation where stmt can be partially true or partially false and provides method to represent uncertainty and vagueness in decision making relies on linguistic variables and fuzzy rules

expert systems

mimic decision making abilities of human expert within specific field draw inferences from specific knowledge base and relies on AI replicate judgment and behavior of human w/ specific expertise widely deployed across industries designed support and assist humans rather than replace

supervised learning models (ML)

most common learn from pre-labeled and classified data set algorithm analyze input data and associated labels to produce inferred function that becomes basis for system to make predictions based on new, previously unseen inputs compare outputs w/ correct/intended output, to ID errors and improve prediction skills 2 subcategories 1) classification modles 2) regression models

4th industrial revolution or industry 4.0

next stage of industry and manufacturing advancements, increased in connectivity and smart automation robotics stems from engineering and computer sicence aims design machines that can perform tasks, usually very specific tasks or duties w/o human intervention AI introduces efficiencies and effective pathways enabling exponential improvement of robotic processes

relevant stakeholders to consider when working w/ AI

ppl who look at broader societal influences of AI such as anthropologists, sociologists or others in social sciences ppl who develop and implement AI systems

AI applications

refer to how system is used examples: -ecommerce -education -healthcare -autonomous vehicles -navigation -FRT -robotics -HR -marketing -social media -chatbots -finance

reinforcement learning models (ML)

repetition based system used reward and punishment matrix to determine correct or optimal outcome rely on trial and error to determine what to do or not and are rewarded or punished accordinly do not ingest pre-labeled datasets, learn solely thru action and repetition, changing or not changing state or getting feedback from environment actions and decisions that result in reward reinforce triggering behavior (incentivize model to follow same tactic in future) errors trigger penalty and reduce reward, proportional in size to scale of error

3 main elements of expert system: inference engine

responsible for extracting relevant info from knowledge base and using it to solve problem uses rule based approach that maps data to series of rules, which system relies on to make decisions in response to input provided expert systems often include module that allow users review decision making process

AI platform

software that allows org to develop, test, deploy and refresh AI apps e.g., Google Cloud, Microsoft Azure, AWS

regression model

subcategory of supervised learning model involve prediction of continuous value, for example estimating stock price e.g., support vector machine (SVM) used for both classification and regression tasks but widely used for classification objectives; support vector regression (SVR) most commonly used produce continuous values

classification model

subcategory of supervised learning models produce outputs in form of specific categorical response e.g., does image contain puppy or not

association rule learning

subcategory of unsupervised learning models ID relationships and associations btwn data points e.g., understanding consumer buying habits

clustering

subcategory of unsupervised learning models automatically group data points that share similar or identical attributes (e.g., DNA samples that share similarities or patterns)

OECD 5 dimensions to classify AI system: tasks and output

tasks that AI systems perform, its outputs and resulting actions from those outputs characteristics include -system tasks -systems that combine tasks and actions -evaluation methods used to look at how tasks and systems perform

common elements of AI

technology - use of tech & specified objectives or tech to achieve autonomy - level of autonomy by teach to achieve defined objectives human involvement - need for human input to train tech & ID objectives for it to follow output - tech produces output, e.g., performing tasks, solving problems, producing content

3 main elements of expert system: knowledge base

typically consist of org collection of facts and info provided by human experts and focused on specific field or domain system also allowed gather additional info from external sources

unsupervised learning

unlabeled data typically used for pattern detection

OECD 5 dimensions to classify AI systems: data and input

what type of data was used in model and any expert input used -expert input = human knowledge codified into rules -includes characteristics such as how data collected and what collection method was used (by machine or by human), structure of data, and data format

OECD 5 dimensions to classify AI systems: ppl & planet

ID ppl and groups that might be affected by AI system e.g., human rights, environment and society privacy comes into play here