EMBODIED NEURO- SYMBOLIC COMPUTATION SERGE THILL - - PowerPoint PPT Presentation

EMBODIED NEURO- SYMBOLIC COMPUTATION

SERGE THILL SERGE.THILL@HIS.SE

CONTENTS

(from stampauctionnetwork.com)

TWO REASONS FOR BEING INTERESTED IN NEURO-SYMBOLIC COMPUTATION

1. IT APPEARS USEFUL FOR APPLICATIONS IN COGNITIVE SYSTEMS 2. IT MAY BE A WAY TO MOVE THE STATE OF THE ART IN EMBODIED THEORIES OF COGNITION FORWARD

COGNITIVE SYSTEMS

• THE MAIN OBJECTIVE OF DREAMS4CAR IS TO DEVELOP A DREAM-

LIKE MECHANISM IN WHICH ROBOTS CAN PRODUCE AN EMULATED WORLD WITH WHICH THEY CAN INTERACT TO DEVELOP AND IMPROVE THEIR PERCEPTION-ACTION SYSTEM.

• THE TRAINED PA SYSTEM WILL THEN BE TRANSFERRED BACK TO THE

“LIVE” SYSTEM, IMPROVING ROBOT SENSORIMOTOR CONTROL IN SITUATIONS THAT HAVE BEEN IMAGINED

APPROACH

• BIO-INSPIRED ARTIFICIAL COGNITIVE

SYSTEMS REPRODUCING THE MAIN BRAIN LOOPS:

1. LAYERED CONTROL ARCHITECTURE GENERATING HIERARCHICALLY- STRUCTURED AFFORDANCES, WITH TOPOGRAPHICAL ACTION SPACE (‘DORSAL STREAM’ AND ‘MOTOR CORTEX’). A SIMPLE VERSION ALREADY IMPLEMENTED IN FP7 ‘ADAPTIVE’ PROJECT. 2. BIAS-ABLE ACTION-SELECTION MECHANISM, ALSO PLAYING AN IMPORTANT ROLE IN ACTION DISCOVERY (’BASAL GANGLIA’) 3. FORWARD EMULATORS, PLAYING AN IMPORTANT ROLE IN DETECTING SALIENT SITUATIONS (‘CEREBELLUM’)

EMBODIED COGNITIVE SCIENCE

COGNITION: THE CLASSIC VIEW

P Perception Action (Illustrations from Wikimedia commons)

THE EMBODIED VIEW

(Adapted from Iwasawa in Pfeifer et al, 2011)

SWISS ROBOTS

(Marinus & Boerkhorst, 1996)

WHAT FLAVOUR WOULD YOU LIKE?

Embodied cognition

Grounded cognition

Situated Cognition

Proper embodiment

Social Cognition Distributed Cognition

Enactivism

Deep embodiment Real embodiment

THE BODY IS OFTEN IGNORED

• “PERHAPS SOMEWHAT SURPRISINGLY, MANY DISCUSSIONS/

NOTIONS OF EMBODIED COGNITION ACTUALLY PAY RELATIVELY LITTLE ATTENTION TO THE NATURE AND THE ROLE OF THE BODY INVOLVED (IF AT ALL)”

TWO ROUTES TO EMBODIED COGNITIVE SCIENCE

Chemero (2009)

SYMBOLS IN EMBODIED COGNITIVE SCIENCE

THE SYMBOL GROUNDING PROBLEM* (HARNAD, 1990)

• (RELATED TO SEARLE’S CHINESE ROOM, BUT NOT THE SAME)
• ARBITRARY, AMODAL SYMBOLS CANNOT CARRY MEANING IN

ANY TRUE SENSE (AS OPPOSED TO IN THE SENSE OF HAVING AN ARBITRARY MEANING ASSIGNED BY AN OUTSIDE OBSERVER)

• SYMBOLS NEED TO BE “GROUNDED” IN THE AGENT’S OWN

EXPERIENCE (LATER PROCESSING DOESN’T NEED TO INVOLVE SENSORIMOTOR SYSTEMS THOUGH)

(*more generally: the representation grounding problem, Dorffner, 1997)

THE DICTIONARY MERRY-GO-ROUND

Push: to press forcefully against in order to move Force: energy or strength Energy: strength of force Strength: the power to resist force La la la la….. Roy (2005)

(MEANWHILE, IN COMPUTATIONAL LINGUISTICS)

• EMBODIED COGSCI RESEARCHERS CLAIM THAT GROUNDED

SYMBOLS ARE NECESSARY TO OVERCOME A NUMBER OF ISSUES (DOVE, 2009), E.G.:

• POLYSEMY (DIFFERENT USES OF THE SAME VERBAL LABEL)
• SYNONYMY (DIFFERENT CONCEPTS USED IN SIMILAR CONTEXTS)
• BUT COMPUTATIONAL LINGUISTICS HAS METHODS TO DEAL

WITH THESE (E.G. AUGMENTING DISTRIBUTIONAL PROFILES WITH RELATIONAL INFORMATION; THILL ET AL., 2014)

THIS DOES NOT RESOLVE SEARLE’S CRITICISM OF STRONG AI

• SOME RESEARCHERS IN “EMBODIED” AI WILL
• AGREE WITH SEARLE’S CRITICISMS OF STRONG AI
• THINK EMBODIED AI SOLVES THE ISSUE

LIKE SO

“SUPPOSE WE PUT A COMPUTER INSIDE A ROBOT, AND THIS COMPUTER WOULD NOT JUST TAKE IN FORMAL SYMBOLS AS INPUT AND GIVE OUT FORMAL SYMBOLS AS OUTPUT, BUT RATHER WOULD ACTUALLY OPERATE THE ROBOT IN SUCH A WAY THAT THE ROBOT DOES SOMETHING VERY MUCH LIKE PERCEIVING, WALKING, MOVING ABOUT, HAMMERING NAILS, EATING DRINKING -- ANYTHING YOU LIKE. THE ROBOT WOULD, FOR EXAMPLE HAVE A CAMERA ATTACHED TO IT THAT ENABLED IT TO 'SEE,' IT WOULD HAVE ARMS AND LEGS THAT ENABLED IT TO 'ACT,' AND ALL OF THIS WOULD BE CONTROLLED BY ITS COMPUTER 'BRAIN.' SUCH A ROBOT WOULD […] HAVE GENUINE UNDERSTANDING AND OTHER MENTAL STATES”--THE SECOND REPLY TO THE CHINESE ROOM ARGUMENT CONSIDERED, AND

REJECTED, BY SEARLE (1980)

LIKE SO

“SUPPOSE WE PUT A COMPUTER INSIDE A ROBOT, AND THIS COMPUTER WOULD NOT JUST TAKE IN FORMAL SYMBOLS AS INPUT AND GIVE OUT FORMAL SYMBOLS AS OUTPUT, BUT RATHER WOULD ACTUALLY OPERATE THE ROBOT IN SUCH A WAY THAT THE ROBOT DOES SOMETHING VERY MUCH LIKE PERCEIVING, WALKING, MOVING ABOUT, HAMMERING NAILS, EATING DRINKING -- ANYTHING YOU LIKE. THE ROBOT WOULD, FOR EXAMPLE HAVE A CAMERA ATTACHED TO IT THAT ENABLED IT TO 'SEE,' IT WOULD HAVE ARMS AND LEGS THAT ENABLED IT TO 'ACT,' AND ALL OF THIS WOULD BE CONTROLLED BY ITS COMPUTER 'BRAIN.' SUCH A ROBOT WOULD […] HAVE GENUINE UNDERSTANDING AND OTHER MENTAL STATES” (SEARLE, 1980, “THE ROBOT REPLY”)-THE SECOND REPLY TO THE CHINESE

ROOM ARGUMENT CONSIDERED, AND REJECTED, BY SEARLE (1980)

SEARLE DISAGREES

• “[T]HE ANSWER TO THE ROBOT REPLY IS THAT THE

ADDITION OF SUCH "PERCEPTUAL" AND "MOTOR" CAPACITIES ADDS NOTHING BY WAY OF UNDERSTANDING, IN PARTICULAR, OR INTENTIONALITY, IN GENERAL, TO SCHANK'S ORIGINAL PROGRAM.”

TWO ROUTES TO EMBODIED COGNITIVE SCIENCE

Chemero (2009) Anti-functionalism (e.g. Searle) Functionalism (e.g. Harnad)

(EMBODIED) NEURO-SYMBOLIC INTEGRATION

NEF AND SPA

• NEF:
• AN ATTEMPT TO ANSWER THE QUESTION “WHAT ARE NEURONS GOOD

AT COMPUTING?”

• SPA:
• AN ATTEMPT TO ANSWER THE QUESTION

“WHAT KIND OF COGNITIVE ARCHITECTURE CAN WE BUILD WITH THAT?

• NENGO:
• A WAY TO BUILD MODELS USING THESE

SEMANTIC POINTERS AND EMBODIMENT

WHAT A SPA MODEL OF HUMAN CONCEPTS MAY LOOK LIKE

WHAT A SPA MODEL OF HUMAN CONCEPTS MAY LOOK LIKE

(Todo: actually implement this J )

PROPER EMBODIMENT

• “[…] RECENT WORK IN NEUROSCIENCE AND ROBOTICS SUGGESTS THAT

COGNITIVE SYSTEMS ARE NOT MERELY SUPERFICIALLY EMBODIED IN THE SENSE THAT THE SENSORIMOTOR INTERACTIONS WITH THE ENVIRONMENT ARE THE ONLY INTERACTIONS RELEVANT TO COGNITIVE BEHAVIOUR, BUT THAT COGNITIVE SYSTEMS ARE “PROPERLY EMBODIED”; THE INTERNAL BODY MATTERS TO COGNITION”. (STAPLETON, 2013)

PROPER EMBODIMENT

• “[...]MANY OF THE COMPONENTS OF OUR PHYSIOLOGY; AFFECTIVE

INFORMATION, GASOTRANSMITTERS, NEUROMODULATORS AND NEUROTRANSMITTERS, HORMONES ETC. ARE NOT MERE BACKGROUND CONDITIONS FOR COGNITIVE PROCESSING BUT ARE AS CONSTITUTIVE* AS THE NEURAL ELECTRICAL PROCESSES ARE.”

EMBODIMENT AT THE NEURAL LEVEL

• “[…] LARGE ASPECTS OF OUR BRAIN AND THE WAY IT PROCESSES DATA

EXIST NOT BECAUSE THEY ARE NECESSARILY NECESSARY FOR INTELLIGENCE BUT BECAUSE OF CONSTRAINTS IMPOSED BY BIOLOGY AND/OR BY OUR PARTICULAR MORPHOLOGY. SOME MAY INDEED BE IRRELEVANT FOR ANYTHING BUT THE HUMAN. OTHERS, HOWEVER, MAY GENERALISE TO OTHER EMBODIMENTS, INCLUDING THAT OF A MACHINE. FINALLY, THERE MAY BE GENERALLY VALID COMPUTATIONAL ASPECTS THAT ARE, IN FACT, INDEPENDENT OF A PARTICULAR EMBODIMENT (INCLUDING MORPHOLOGY AND BIOLOGY)”

CAVEAT: ASSUMPTIONS ABOUT NEURAL COMPUTATIONS SHAPE THEORIES OF COGNITION

• “DNS CAN SERVE AS A GOOD MODEL OF BEHAVIOUR ONLY IF WE

PRETEND THAT BEHAVIOUR AMOUNTS TO ONE GOOGLE (OR WATSON) QUERY AFTER ANOTHER.”

CAVEAT: ASSUMPTIONS ABOUT NEURAL COMPUTATIONS SHAPE THEORIES OF COGNITION

• “INSOFAR AS OUR HYPOTHESES REGARDING THE PROBLEMS THAT ANIMALS

OSTENSIBLY SOLVE ARE BASED ON ABDUCTIVE REASONING, THEY ARE SUBJECT TO

• BIASES. SOME OF THESE BIASES ARE ENTIRELY OF OUR OWN MAKING. FOR INSTANCE,

IF AN EFFECTIVE ALGORITHMIC FRAMEWORK HAPPENS TO BECOME AVAILABLE FOR A CLASS OF PROBLEMS, THESE THEN TEND TO LOOM LARGER AS POSSIBLE FACTORS IN THEORIES OF BEHAVIOUR. AS A QUICK EXAMPLE (TO BE REVISITED LATER), CONSIDER THE WIDESPREAD INFLUENCE EXERTED ON COGNITIVE SCIENCE BY THE INVENTION OF THE BACK- PROPAGATION ALGORITHM FOR TRAINING MULTILAYER PERCEPTRONS”

[FORMAL LANGUAGES FOR MODELLING COGNITION]

ON MATHEMATICS AS A LANGUAGE FOR THEORIES AND MODELS

“THE MIRACLE OF THE APPROPRIATENESS OF THE LANGUAGE OF MATHEMATICS FOR THE FORMULATION OF THE LAWS OF PHYSICS IS A WONDERFUL GIFT WHICH WE NEITHER UNDERSTAND NOR DESERVE”

• - EUGENE WIGNER, 1960

ON MATHEMATICS AS A LANGUAGE FOR THEORIES AND MODELS

“MATHEMATICS IS THE SCIENCE OF SKILLFUL OPERATIONS WITH CONCEPTS AND RULES INVENTED JUST FOR THIS PURPOSE.” “MOST MORE ADVANCED MATHEMATICAL CONCEPTS […] WERE SO DEVISED THAT THEY ARE APT SUBJECTS ON WHICH THE MATHEMATICIAN CAN DEMONSTRATE HIS INGENUITY AND SENSE OF FORMAL BEAUTY. IN FACT, THE DEFINITION OF THESE CONCEPTS, WITH A REALIZATION THAT INTERESTING AND INGENIOUS CONSIDERATIONS COULD BE APPLIED TO THEM, IS THE FIRST DEMONSTRATION OF THE INGENIOUSNESS OF THE MATHEMATICIAN WHO DEFINES THEM”

• - EUGENE WIGNER, 1960

“MATHEMATICS DEALS EXCLUSIVELY WITH THE RELATION OF CONCEPTS TO EACH OTHER WITHOUT CONSIDERATION OF THEIR RELATION TO EXPERIENCE”

• - ALBERT EINSTEIN, THE INTERNET

ON MATHEMATICS AS A LANGUAGE FOR THEORIES AND MODELS

• YET MATHEMATICS CONTINUOUSLY PROVES ITSELF EXTRAORDINARILY USEFUL IN THE

EXPLANATION OF NATURAL PHENOMENA IN WAYS THAT CAN NOT REASONABLY BE EXPECTED

• EXAMPLE (SEE WIGNER, 1960, FOR MORE):
• SOME RULES OF COMPUTATION GIVEN BY HEISENBERG APPEARED FORMALLY EQUIVALENT TO

MATRIX COMPUTATIONS (ESTABLISHED LONG BEFORE HEISENBERG).

• IT THEN TURNS OUT THAT MATRIX MECHANICS ARE ALSO APPLICABLE UNDER MORE REALISTIC

CONDITIONS (HEISENBERG USED THE “OLD” THEORY OF THE ATOM, USING THE CLASSICAL NOTIONS OF MOTION)

• THE LOWEST ENERGY LEVEL OF HELIUM WAS THUS CORRECTLY (TO WITHIN 1 PART IN 1

MILLION) CALCULATED.

• HOW EXACTLY DID A MATHEMATICAL CONCEPT WHICH HAS, IN ITSELF, NOTHING TO DO

WITH HELIUM ENERGY LEVELS, TURN OUT TO BE SO ACCURATE?

OBLIGATORY XKCD

(from xkcd.com)

THANKS

J