知能と環境 Intelligence and Environment. 中島秀之 Hideyuki Nakashima 公立はこだて未来大学 Future University Hakodate

Transcription

1 知能と環境 Intelligence and Environment 中島秀之 Hideyuki Nakashima 公立はこだて未来大学 Future University Hakodate

2 HISTORY OF AI H. Nakashima 4

3 History of Frameworks of Intelligence 1. Physical Symbol System Hypothesis:The essence of intelligence is the formal symbol manipulation (by the founder of AI field - Newell, Simon, Minsky, McCarthy, etc.) 2. The essence of intelligence is in Pattern recognition (articulation of the world) Neural networks, vision Deep learning 3. The essence is in the interaction with the environment Subsumption architecture (Brooks 1991) Autopoiesis (Maturana and Varela 1980) Situatedness (situated cognition, situation theory) 2016/5/12 H. Nakashima 5

4 Meaning/Role of Symbols Articulation of the world Extraction of equivalence classes by meaning to intelligent entities The same reactions to the same class situations Externalization is not the necessary condition Some of middle layer nodes of NN function as symbols Externalize when used for communication Symbol grounding problem is how to connect symbols to existence in the real world in a intrinsic way 2016/3/3

5 Frame Problem Problems in symbolic representation of actions Original version (McCarthy and Hayes 1969) Explosion of description Explosion of processing (inference) Extended version (Hanks and McDermott, "Yale Shooting Problem, 1986) Qualification problem Impossibility of describing all pre-conditions of actions Ramification problem Impossibility of inferring all outcomes of actions 2016/3/3

6 Symbol Grounding Problem (Harnad 1990) How a symbol is grounded into the real world object(s)/phenomenon Only a robot can do it? Isolated machine cannot do it A robot has its own experience (interaction with the world) It is still unknown if a robot can get the meaning of words similar to and communicable with human 2016/3/3

7 View of Intelligence in the Initial Stage of AI: Physical Symbol System Hypothesis Intelligent system action reasoning recognition Physical symbol system = Formal system Reasoning with internal representation only Frame Problem was identified environment 2016/5/12 H. Nakashima 9

8 Subsumption Architecture Intelligent system recognition reasoning action Brooks Intelligence of insects From horizontal (sequential) flow to vertical (parallel) flow Higher layers subsume lower layers environment 2016/5/12 H. Nakashima 10

9 Enhancement of the Interaction with the Environment Intelligent system subject recognition reasoning action Uexküll: Umwelt (cf. Umgebung) Gibson: affordance Maturana & Varela: Autopoiesis Situatedness environment 2016/5/12 H. Nakashima 11

10 Research around Umwelt is the perceptual world in which an organism exists and acts as a subject. cf. Umgebung Uexküll: Umwelt (1934) 2016/3/3

11 2016/3/3 Uexküll Wirkkreis (Effective Circle)

12 Umwelt of Hermit Crab encounter with sea anemone With shell camouflage Without shell shell Hungry food 2016/3/3

13 Robot is Intelligence with Actuators Christmas Bush (designed by Hans Moravec)

14 FUTURE

15 Problems of Machine Learning Over fitting Easy to create false positive example (over generalization) Human also has illusion Rare Close pattern only Use of top-down expectation may (should) solve those problems 2016/3/3

16 Fusion of Top-down Expectation and Bottom-up Input Target: Realization of "Constructive Intelligence " Hierarchical regions of the brain (cerebral cortex) predict their future input sequences (top-down process) Deep learning works basically bottomup (no expectation or top-down guidance) Human recognition is said to be: Mainly top-down Key to solve machine learning problems Bottom-up inputs just trigger them 2016/3/3

17 Noema and Noesis (by Bin Kimura) Future Noema: Image of music C1:Play C 2 : Interaction with the environment Conceptual world C3:Plan Current noema: music played C2:Comprehension Produced sound Physical world A bit different from Husserl s use Example: playing music C1:action of the player C1 2:produced sound, echo, audience reaction, etc. C2:Listening to the music C3:Planning the next movement 2016/3/3

18 FNS Diagram:Model of Synthetic Methodology Goal C1:Generation C 2 : Interaction with the environment Conceptual world C3:scripting Property C2:Analysis Generated entity Physical world Future Noema Synthesis Spiral because goal will change Fractal because each arrow may be expanded to another FNS C2 in particular is ANALYTIC LOOP 2016/5/12 H. Nakashima 21

19 Model of Constructive Intelligence Intension C1:Action C 2 : Interaction with the environment Conceptual world C3:Orientation Recognition C2:Observation Behavior Physical world Close to Kimura model Intelligence capable of producing desired future Action (C1):top-down (with feedback through C2, C3) Observation (C2):bottomup (with top-down control) Orientation (C3):thought process, symbol manipulation 2016/3/3

20 Cognition and Reality (Neisser 1976) C3 expectation recognition C1 C2 C 2 :interaction with the environment