WIXLIB

tonomousGeneralAutonomous GeneralK. R. Thórisson 2021CumulativeCu m u l a t i v e LearningLearningKristinn R. ThórissonP R O F E S S O R , R E Y K J A V I K U N I V E R S I T Y, D E P A R T M E N T O F C O M P U T E R S C I E N C EDIRECTOR, ICELANDIC INSTITUTE FOR INTELLIGENT MACHINEScadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Overview Intelligence Cumulative learning Causal relations & reasoningK. R. Thórisson 2021 Seed-programming, autonomy & generalitycadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

General Machine IntelligenceIntelligence: Controller body thatK. R. Thórisson 2021 achieves complex goals in novel task-environments processing capacity available infocadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

K. R. Thórisson 2021Agent, Body, Controllercadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Agent, Body, ControllerenvironmentK. R. Thórisson 2021agentcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Agent, Body, blesK. R. Thórisson 2021manipulatablevariablescadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Agent, Body, lermodelingprocessesK. R. Thórisson 2021manipulatablevariablescadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Agent, Body, blesK. R. Thórisson 2021manipulatablevariablescadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Agent, Body, Controllertask-environmentagent bodyvariablesobservablevariablesK. R. Thórisson 2021manipulatablevariablescadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

General Machine IntelligenceIntelligence: Controller body thatK. R. Thórisson 2021 achieves complex goals in novel task-environments processing capacity available infocadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

General Machine IntelligenceIntelligence: Controller body thatK. R. Thórisson 2021 achieves complex goals in novel tasks & environments processing capacity available infocadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Complex Task-Environment Large number of variables, relations, andtransformation functions Complex spatio-temporal patternsK. R. Thórisson 2021 Novelty is commoncadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Complex Task-EnvironmentGiant Large number of variables, relations, andtransformation functions Complex spatio-temporal patternsK. R. Thórisson 2021 Novelty is commoncadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Complex Task-EnvironmentGiant Large number of variables, relations, andtransformation functions Complex spatio-temporal patternsK. R. Thórisson 2021 Novelty is commoncadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Complex Task-EnvironmentGiant Large number of variables, relations, andtransformation functions Complex spatio-temporal patternsK. R. Thórisson 2021 Novelty is commonthe rulecadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

K. R. Thórisson 2021What is Intelligence?cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

What is Intelligence?K. R. Thórisson 2021“Figuring out how to get new stuff done”cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

What is Intelligence?new you haven’t done it before“Figuring out how to get new stuff done”K. R. Thórisson 2021Figuring out new some missing information AIKR(assumption of insufficientknowledge and resources)cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

What is Intelligence?new you haven’t done it before“Figuring out how to get new stuff done”Getting stuff done make something happen AER(assumption of the existenceof regularity)K. R. Thórisson 2021Figuring out new some missing information AIKR(assumption of insufficientknowledge and resources)cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

What is Intelligence?new you haven’t done it before“Figuring out how to get new stuff done”K. R. Thórisson 2021Figuring out new some missing information AIKR(assumption of insufficientknowledge and resources)Getting stuff done make something happen AER(assumption of the existenceof regularity)Why non-axiomatic reasoning?cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

K. R. Thórisson 2021Why Non-Axiomatic Reasoning?cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Why Non-Axiomatic Reasoning?K. R. Thórisson 2021For any world with regularitythe onlyeffective and efficient methodfor handlingcomplex task-environments is viasystematic application of logic- i.e. reasoning.cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Why Non-Axiomatic Reasoning?For any world with regularitythe onlyeffective and efficient methodfor handlingcomplex task-environments is viasystematic application of logic- i.e. reasoning.:K. R. Thórisson sVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Why Non-Axiomatic Reasoning?Because when you do NEW stuffyou cannot possibly know if it will work:You don’t know any axioms(or whether any exist).For any world with regularitythe onlyeffective and efficient methodfor handlingcomplex task-environments is viasystematic application of logic- i.e. reasoning.:K. R. Thórisson sVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Why Non-Axiomatic Reasoning?K. R. Thórisson 2021How does that proceed?cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Why Non-Axiomatic Reasoning?How does that proceed?K. R. Thórisson 2021 . through hypothesis generationand testing,i.e. modeling.cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Why Non-Axiomatic Reasoning?How does that proceed?K. R. Thórisson 2021 . through hypothesis generationand testing,i.e. modeling.learning modelingcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Why Non-Axiomatic Reasoning?How does that proceed?K. R. Thórisson 2021 . through hypothesis generationand testing,i.e. modeling.learning modeling reasoningcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Cumulative Learning What kind of information structuresare needed for a controller to modelcomplex environments?cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

K. R. Thórisson 2021CAUSATIONcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

CAUSATIONK. R. Thórisson 2021- the basis forgeting anything donecadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

CAUSATIONK. R. Thórisson 2021- the basis forgeting anything done- you can’t get anything doneif you don’t know what leads to whatcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Why We Need Causal RelationsK. R. Thórisson 2021 In a world containing regularity, composed of unknownphenomena a small subset of which are manipulatable, and a small subset of which are observable you need to know how they relate: which variables affect which other variables, and how toknow that they do. We say that variables with relations that allow us to getstuff done are causally connectedcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Causal-Relational Models CRMs are models of relationsK. R. Thórisson 2021 i.e. information structures that describe actionable predictions created based on experience capture relations between observed patterns Model relations between variables andelements in Eα t— βt1cadia.ru.isVISCA-2021 LECTURE SERIES 1 dUNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Causal-Relational Models The process of creating and using modelsinvolves testing them to evaluate theirusefulness:K. R. Thórisson 2021 the more accurately they help a controllerachieve goals the more useful they arecadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Causal-Relational Models To evaluate models’ usefulness, they must be falsifiable, just like hypotheses in an empiricalcomparative experiment. This means their creation must be boundedby practical concernsK. R. Thórisson 2021 e.g. by limiting new models primarily toobservable patterns and variables.cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Generating CRMsK. R. Thórisson 2021α, β, γ : observable variablesActualphenomenonin thephysical worldα βα γcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Generating CRMsMODELSM1 : αM2 : αM3 : γβγβM4 : βγK. R. Thórisson 2021α, β, γ : observable variablesActualphenomenonin thephysical worldα βα γcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Generating CRMsK. R. Thórisson 2021MODELSM1 : αM2 : αM3 : γβγβM4 : βγ Any of these will predict observed events correctly: If you see β you will see γ, and vice versa If you see α you will see β and γcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Generating CRMsK. R. Thórisson 2021MODELSM1 : αM2 : αM3 : γβγβM4 : βγ However, if you want to stop seeing γ it does not help to makeβ go away, or vice versa To achieve a sub-goal of removing β (or γ) the only variable thatcan achieve that is α, captured only by M1 (or M2)cadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

CRMs are Bi-Directional CRMs are bi-directional in that read “forward” (left to right) they meanthat A may cause BK. R. Thórisson 2021 read “backward” they mean that B mayhave been caused by Acadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

CRMs are Bi-Directional CRMs are bi-directional in that read “forward” (left toIOright)N they meanTCthat A may causeBUDEDK. R. Thórisson 2021 read “backward” they mean that B mayhave been caused by Acadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

CRMs are Bi-Directional CRMs are bi-directional in that read “forward” (left toIOright)N they meanTCthat A may causeBUDEDK. R. Thórisson 2021 read “backward” they mean that B mayN by AOhave been causedITCUDABcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

CRMs are Bi-Directional CRMs are bi-directional in thatNmean read “forward” (left toIOright)N theyOITTCCIthat A may causeUBDEDERDPK. R. Thórisson 2021 read “backward” they mean that B mayN by AOhave been causedITCUDABcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

CRMs are Bi-Directional CRMs are bi-directional in thatNmean read “forward” (left toIOright)N theyOITTCCIthat A may causeUBDEDERDPK. R. Thórisson 2021 read “backward” they mean that B mayN by A INGOhave been causedINTNCULADPABcadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is

Causal KnowledgeK. R. Thórisson 2021Like comparative empirical experimentscadia.ru.isVISCA-2021 LECTURE SERIES UNIVERSITY OF MICHIGAN JUNE 8 2021 ZOOM.COMwww.iiim.is