human-level ai & phenomenology
TRANSCRIPT
2015-07 12015-07
ARAKAWA, Naoya, Ph.D
Human-Level AI & Phenomenology
2015-07-11
2015-07 2
Today’s Topic
● Creating Human-Like AI
○ Background, Issues & Approaches
○ Its relation to Embodiment &
Phenomenology
○ My recent activities
2015-07 3
Abridged CV
• Education–Undergraduate : Brain & Neural Nets
–Graduate (M.E.) : Systems Science
–Ph.D : Philosophy of Language“The Naturalization of Reference”
• Work: Natural Language Processing
– Machine Translation
– Dialog systems
– Semantic analysis, Ontology compiling
• Recent activities: Artificial General Intelligence
2015-07 4
Table of Contents
1.Background for Human-Level AI● AI with Human cognitive functions● Recent ‘AI boom’● Two contrapositions
2. Issues to be solved
3. How to create Human-Level AI
4. My recent activities
2015-07 5
Human-Like AI
● An aim/ambition of the AI discipline
○ 「 Agalmatophilia 」 ?
○ AI as 「 Cognitive Science 」
● Constructive (Make & Test) Understanding of
Human-beings
○ Build to understand
○ Difficulty in fully analytic understanding
2015-07 6
Recent “AI Boom”
● Media Coverage○ AI books for general public
○ TV programs on AI
○ New research centers
● Technological Background○ Computing Power○ Availability of “Big Data”○ Some notable results: Chess, Jeopardy!, Self-driving
cars, ...○ Advances in Machine Learning
Deep Learning! ⇒
2015-07 7
Advances in Machine Learning
● The Neural Net Strikes Back!
● Deep Learning○ Multi-Layered Neural Networks
○ Notable results in pattern recognition
○ Automatic concept formationGoogle Brain (Cat), Google Dreams (Inceptionism)
● Recurrent Neural Network (RNN)○ Learning time-series○ Captioning images with deep learning (Stanford U.)
● Reinforcement Learning○ Learning action sequences based on rewards
○ Deep Q Network: playing Atari games
2015-07 8
AGI vs. Narrow AI
● Artificial General Intelligence vs. Narrow AI○ Artificial General Intelligence
■ ‘General’ in the sense that it can learn various skills■ Human-Like AI AGI⊂■ Long hoped... but difficult to realize⇒
○ Narrow AI: to solve specific issues〜 the current main stream
● GOFAI vs. Emergentist AI○ Good Old-Fashined (Symbolic) AI
■ Criticized by thinkers such as Dreyfus & Lakoff■ Knowledge acquisition bottleneck
○ Emergentist AI■ Knowledge is not to be given but to learn■ Analog (statistic)※Advances in machine learning AGI sees the light here!?⇒
2015-07 9
Table of Contents
1.Background for Human-Level AI
2. Issues to be solved ● Knowledge Acquisition = Learning = Epistemology● Cognitive Functions
2. How to create Human-Level AI
3. My recent activities
2015-07 10
Issues to be solved
Knowledge Acquisition = Learning= Epistemology● How do we get knowledge?
● How do machines get knowledge?
● More concretely:
○ Acquistion of concepts ←( perception & motion )
○ Knowledge acquisition on action( praxis/pragmatics←motion & perception )
○ Language Acquistion■ Acquistion of Vocabulary (the Symbol Grounding Problem)
■ Acquistion of Grammar
2015-07 11
Cognitive Functions to be realized
○ Human-Level AI⇔Inventory of Human Cognitive Functions○ Learning 〜 Knowledge Acquisition
■ Pattern Recognition (mostly supervised)■ Conceptual Learning (mostly unsupervised)
● ‘Clustering’● ‘Representation Learning’ in Deep Learning
■ Reinforcement Learning : learning action sequences based on rewards■ Episodic Memory : One-shot Learning
○ Planning & Execution■ Emergentist AI: trying to get inspiration from the prefrontal cortex?
○ Linguistic Functions■ Generativity ( Syntax )■ Social aspects ( Pragmatics )■ Grounding ( Semantics )
2015-07 12
Table of Contents
1.Background for Human-LevelAI
2. Issues to be solved
3. How to create Human-Level AI● Three Pillars● Make & Test (Constructive) Approach
2. My recent activities
2015-07 13
How to Create Human-Level AI
1.Three Pillars ( IMHO )•Cognitive Architecture: Overall Structural ModelsIntelligence has ‘structure’Traditional ones: symbolicYou can learn from the brain too.
•Machine LearningMathematical models for learning
•Cognitive Robotics (embodiment)Learning developmentally in the environment
2.The Constructive (Make & Test) Approach• Hypotheses robots/simulation to corroborate⇒• Cognitive Robotics• Artificial Brains
2015-07 14
Cognitive Robotics
• Robotics as Cognitive Science• Stance: cognition requires the body.• ‘Constructive’ understanding of cognition
Construct to understand!
• Genres– Cognitive Developmental Robotics
• Developing cognitive abilities like human children
– Robotics for Symbol Emergence• Learning language via interaction with the environment
– Robotics for Social Intelligence• Communicating robots
2015-07 15
Cognitive Developmental Robotics
• Developing cognitive abilities like human children• Robots learns from interaction with the
environment• To complement experiments with human infants
(which are difficult for ethical reasons)• Researches in Japan, e.g.:
–Asada Lab. @ Osaka U.–Kuniyoshi Lab. @ Tokyo U.–The Constructive Developmental Science @ MEXT
• Ref.– Cangelosi, A. et al.: Developmental Robotics
-- From Babies to Robots, MIT Press (2015).– Asada M. et al.: "Cognitive developmental robotics: a survey," in IEEE Transactions
on Autonomous Mental Development, Vol.1, No.1, pp.12--34 (2009)
2015-07 16
Robotics for Symbol Emergence
• Learning language via interaction with the environment• Human-beings : no grammar, no vocabulary given• ref. Developmental Linguistics
– Tomasello, Meltzoff, Spelke, …– Chomskians ( the merge theory )– cf. Evolutional Linguistics (animal cognitive functions)
• The Symbol Grounding Problem: mapping symbols to things in the world
• Machine learning methods– Non-parametiric bayes, Recursive Neural Net…
• Getting insights from developmental linguistics• Yet to succeed in language acquistion
2015-07 17
Robotics for Social Intelligence
● Communicatin study with robots
● Communication requiring the body
● Mimetics
● Joint attention
● Empathy
2015-07 18
Cognitive Robotics & Embodiment
• The interests of cognitive robotics researchers〜 the interests of embodiment researchers
• Common terms– Body Image & Body Scheme, etc.
2015-07 19
Artificial Brains
● Reproducing human cognitive functions by creating something similar to the brain
● Brain Simulation○ Physiological models
○ Blue Brain Project, Neurogrid Project, etc.
● Brain-Inspired Cognitive Architectures○ Examples
■Nengo/SPAUN (C. Eliasmith et al.)
■Leabra (O’Reilly et al.)
■The Whole Brain Architecture (to be mentioned later)
2015-07 20
脳研究の現状
● Advance in functional brain imaging (e.g., fMRI)● Cognitive Neuro-Scientists
○ A. Damasio : Somatic Marker Hypothesis ( role of emotion )
○ V.S. Ramachandran : presenting cognitive disorders○ E. Kandel : memory research
○ E. Goldberg : cerebral hemispheres & prefrontal cortex
● Modeling cerebral organs○ Cerebral cortex & areas ( perception, motion, planning, …)
the uniform structure of cortex [Mountcastle]
○ Basal ganglia (striatum, etc.: reinforcement learning, WM…)
○ Limbic System (amygdala, etc.: emotion, reward,...)
○ Hypocampus (memory, space representation)
○ Cerebellum (motion control, higher-order cognitive functions)
⇒ To draw an integrated picture soon?
2015-07 21
The Brain and Cognitive Functions ( Figure )
Prefrontal Cortex: Planning
Motor Area : Motion Sequences
Basal Ganglia :Reinforcement Learning
Cerebellum : Feed-forward prediction ?
Hypocampus : Episodic Memory( Place Memory in Rodents )
Where Path
What Path
Amigdalae, etc. :Emotion
Language Areas
To think of an ‘architecture’ constituting of such functional modules to realize human-level intelligence
2015-07 22
Table of Contents
1.Background for Human-LevelAI
2. Issues to be solved
3. How to create Human-Level AI
4. My recent activities● Issue of Semantics● Overall Objectives● Phenomenology of Artefacts ( Manifesto )● Phenomenology of Time● Language Acquistion by Artifacts● AGI related activities
2015-07 23
Semantic Issue : doubts from my pre-history
• Creating an ontology for natural language• The problem of polysemy (ambiguity)
– How many senses?E.g., prepositions
– Border-line uses...
• How do humans acquire word senses?
• Keys in human developmental process
• Counsel by Lakoff, the Cognitive Linguists
Women, Fire, and Dangerous Things
It is impossible to deal with meaning with symbolic logic!
⇒ Radical readdressing is required!
2015-07 24
Overall Goal : Explaining Cognition
● More precisely : Grounding Semantics● But semantics requires epistemology.
○ No sense made without knowing the world.● By-product : AGI/Human-Leval AI
○ But the by-product is the mean in the constructive method.
⇒ Methodological Loop
2015-07 25
Approach
● Learning from animals○ Modeling brains, comparative psychology, etc.
● Phenomenological & Developmental○ Knowledge acquisition from information given to
individuals ● Constructive (make & test)
○ Machine Learning
○ Robotics ( simulation )● Language Acquistion
○ Language : an essential component of cognition
○ Explanation with 1 〜 3 above
2015-07 26
Phenomenology of Artefact (2014-02)
• Husserlean phenomenology 〜 Grounding Epistemology• Epistemology from the first person view• Robots has the first person view
Video : MIT Atlas robot - first person view sensor visualization ⇔
• Robots with kinesthetics • Developmental knowledge acquistion• Information processing with robots
– inspectable– methematically verifiable
• Time consciousness with machine learning? ⇒ Reconstructing phenomenology with artifacts (robots)?
2015-07 27
Phenomenology of Time
● Time Consciousness by Husserl: Urimpression, Protention, Retention
● Time-series Learning 〜 Time-series Prediction○ RNN (recurrent neural network)
○ Temporal Cerebral Models : HTM, DeSTIN, etc. ( cf. akinestopsia @V5 )○ PSI model by Dörner (cognitive psychologist)
Bach J.: Principles of Synthetic Intelligence -- PSI: An Architecture of Motivated Cognition , Oxford U.○ LLoyd, M.: “Time after Time -- Temporality in the dynamic brain,” Being Time: Dynamical Models of Phonomenal
Experience, John Benjamins Pub. Co. (2012)
● Time-series Learning & Phenomenology of Time○ Protention : memory of the future (prediction)
○ Retention : memory of the context (the internal state from the past input)
○ Urimpression⇔ contextualized (differential) present
● cf. Jun Tani, the roboticist○ RNN
○ Ref. to Husserlean phenomenology of time: longitudinal/transverse intentionality
2015-07 28
Towards Language Acquistion by Artifacts
• Developmental Robotics in the virtual world
• Learning from Infants’ language acquistion•Spelke
•Concepts of things: certain constraints–cf. Quine: “Gavagai”
–Seeing thing as a whole
cf. Husserl: looking around objects 3D object concept⇒
•Tomasello• Understanding reference by others requires understanding intention.
•Usage-based grammar learning (anti-generative grammar)
•Meltzoff
•Infants’ understanding of the intention of others
•Modeling own intentional motions first?
2015-07 29
Towards Language Acquistion by Artifacts (cont.)
• Acquistion of Verbs•Verbs are the crux of sentence structure•Acquired after object/nominal concepts•Modeling own intentional motions first (←Meltzoff)?
cf. sense of agency
Own intention is ‘given’
•Mapping to verbs• ‘Parental’ verb uses
•Pragmatic success/failure of own utterances
• Acquistion of syntax• Concatenating subsequent structures Merge?⇒
• Language acquistion with machine learning
2015-07 30
AGI-related Activities ( ads :- )
❖Dwango AI Lab.● Brain/Cognitive Modeling, Language Acquistion, etc.
❖The Whole Brain Architecture Initiative (NPO)● Brain-inspired cognitive architecture
● Education, promotion
❖SIG AGI ( @ Japanese AI Society )● a reading group
● planning to publish a textbook (in Japanese)…
❖Web site in Japanese●www.sig-agi.org
●Facebook Group
For more information, contact [email protected]