+ All Categories
Home > Technology > Symbol emergence in robotics @ Shonan meeting 2013/11/13

Symbol emergence in robotics @ Shonan meeting 2013/11/13

Date post: 09-May-2015
Category:
Upload: tadahiro-taniguchi
View: 313 times
Download: 1 times
Share this document with a friend
25
Tadahiro Taniguchi, Associate professor College of information tech.&sci., Ritsumeikan univ. Shonan village Shonan meeting 2013/11/11 14
Transcript
Page 1: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Tadahiro Taniguchi, Associate professorCollege of information tech.&sci., Ritsumeikan univ.

Shonanvillage

Shonan meeting2013/11/11 ‐ 14

Page 2: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Tadahiro Taniguchi @tanichu Associate professor, Emergent system laboratory College of information science and technology, 

Ritsumeikan University, Japan 2006, PhD Eng. , Kyoto university  2008‐ Assistant professor 2010‐ Associate professor 

Machine  learning

Cognitive robotics

Decentralized autonomous system

Human communication

Symbol Emergence in Robotics

Page 3: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Computational understandings of mental development‐ From behavioral learning to language acquisition ‐

A human child acquires many physical skills, concepts and knowledge including language through physical and social interaction with his/her environment.

How can we become able to communicate using symbolic system?

We’d like to obtain computationalunder standings of  mental development and symbol emergence.

3

Constructive approach towards  human intelligenceenabling semiotic communication

Page 4: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Symbol grounding problem How can a robot “ground” his/her symbol to the real physical world?  [Harnad ‘90] The robot has to give some meaning to a symbol in its symbolic system designed by a human designer  through sensor‐motor interaction with its environment.

Arbitrary nature of symbol (in semiotics) Arbitrariness of labeling/naming Arbitrariness of categorization/segmentation

SGP implicitly assumes that human “arbitrarily evolved” symbolic system is a “true” symbolic system.

Should a symbolic system for a robot be same as human’s ?

SGP ismissing

Page 5: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Understanding that symbolic system is an emergent property of human cognitive and social system “Worlds of robots”

“Worlds of animals”  (Uexküll 1934) Umwelt (self‐centered world) Animals can receive information only from 

their sensor‐motor system. A human has to obtain various behaviors, concepts 

and language on the basis of experiences in his/her umwelt (closed cognitive system).

Human Symbolic system should be understand on the basis of human embodiment (sensor‐motor system). 

Concepts have to be formedon the basis of sensor‐motor information 

in a bottom‐up way

Ernst Mach’s famous picture

"Early Scheme for a circular Feedback Circle by Uxkull

Page 6: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Sharedsymbolic system

Social constraints in semiotic communication

Concept formation is not “enough” for semiotic communication.

We have to share a symbolic system involving syntax, semantics and pragmatics lexicon and mutual belief

Sheep the  coming,aren’t they!!!

(true) A thief  is coming!! 

bad naming

bad syntax

bad interpretation

It’s impossible to help her

Page 7: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Sharedsymbolic system

Social constraints in semiotic communication

Concept formation is not “enough” for semiotic communication.

We have to share a symbolic system involving syntax, semantics and lexicon pragmatics and mutual belief

A thief  is coming!! 

!

constraintconstraint

constraint

constraint

speech generation

situation recognitioninterpretation

decision making(;O;)

Page 8: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Emergent System Emergent property

Through local interaction between elements of a system, global (or macro‐level) order or pattern emerges. The macro‐level order becomes constraints of micro‐level dynamics, and governs the micro‐level interaction. Such bidirectional process makes the system have novel function, morphology and/or behavior. 

constraintemergencemicro‐macro loop

Before introducing “symbol emergent system”, please remind....

Page 9: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Symbol emergent system

Shared lexicon, syntax,pragmatics, semantics,and mutual belief

Concept formation throughinteraction with environment

Semiotic / Social interaction

Physical interaction

Symbolic SystemEmergence

谷口忠大. 「コミュニケーションするロボットは創れるか‐記号創発システムへの構成論的アプローチ」(2010).Tadahiro Taniguchi “Constructive approach toward symbol emergence system” (in Japanese)

Page 10: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Symbol emergence in robotics (SER) Constructive approach toward symbol emergent systems “Constructive” viewpoint to an intelligent system which uses symbolic system.

1. Constructive approach Understanding human intelligence by constructing robots obtaining symbolic system

2. Constructivism  Understanding human intelligence which construct his/her subjective world (Umwelt).

(c) Nagai lab.

Understandingby developing “models”

Page 11: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Topics in SER

multimodalcommunication

conceptformation

language acquisitionand mental development

learning motor skillsand segmentation of

time‐series infomation

emergence of communication

learning interaction strategy

Prof. Nagai

Page 12: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Segmentation of sensor‐motortime‐series information Dynamics in sensor‐motor time‐

series information changes depending on its environmental situations.

An autonomous system should differentiate such situations and obtain representations.

How do humans and how can the robots segment the time‐series data?

What is the criteria of segmentation?What is an adequate computational algorithm for the segmentation?

Page 13: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Modular learning architecturefor segmenting sensor‐motor time series MOSAIC [Kawato et al.]

Mixture of experts [Jacobs et al.] HAMMER [Yiannis et al.]

Dual schemata model [Taniguchi et al.]

http://www.cns.atr.jp/cnb/HarunoG/harunoG.ja.html

How can we grasp long‐term context? 

“Segment” corresponds toa linear systemlocal informationa short‐term event

T. Taniguchi, T. Sawaragi, "Incremental acquisition of multiple nonlinear forward models based on differentiation process of schema model“, Neural Networks, Vol.21 (1), pp.13‐27 .(2008)

Page 14: Symbol emergence in robotics @ Shonan meeting 2013/11/13

階層的な分節構造の自己組織的な獲得

Representation of a room  and a part of a room and was encoded  in RNNin a hierarchical manner.

Hierarchical mixture ofRNN experts [Tani and Norfi 1999]

Dynamical System ApproachDynamical System Approach Symbols are implicitly obtained. It is difficult to understand how the system formed concepts.

Page 15: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Double articulation structure insemiotic data Semiotic time‐series data often has double articulation

Speech signal is a continuous and high‐dimensional time‐series. Spoken sentence is considered as a sequence of phoneme People don’t give a meaning to a phoneme, but give a meaning to a wordwhich is a sequence of phoneme.

h a u m ʌ́ tʃ I z ð í s

[h a u ] [m ʌ́ tʃ] [ i z ] [ð í s]

How much is this?semantic

(meaningful)

meaningless

unsegmented

Page 16: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Double Articulation Analyzerto estimate latent double articulation structure

Unsupervised learning Estimating 

Language model Emission distribution Segments and chunks

Conditions Unknown number of 

words and letters Unknown emission 

distributions Inference

Approximate Inference Procedure of Double Articulation Analyzer [Taniguchi ‘11]

Nonparametric Bayesian  approach

iHMM[Fox ‘07]

NPYLM[Moachihashi ‘09]

Tadahiro Taniguchi, Shogo Nagasaka, Double Articulation Analyzer for Unsegmented Human Motion using Pitman‐Yor Language model and Infinite Hidden Markov Model, 2011 IEEE/SICE SII.(2011) 

Page 17: Symbol emergence in robotics @ Shonan meeting 2013/11/13

sticky HDP‐HMM [Fox ‘07]

An infinite HMM is an HMM which can estimate the number of hidden state flexibly (potentially infinite)[Beal ‘02] [Teh ‘06]

Sticky HDP‐HMM is a generative model for iHMMwith a stickiness parameter [Fox ‘07]. 

Fox et al. developed fast inference algorithm with weak‐limit approximation.

Gaussian emission distribution

β

z1

γ

λ θk∞

πkα

y1

z2

y2

z3

y3

zT

yT

κ

iHMM[Fox ‘07]

E.B. Fox, E.B. Sudderth, M.I. Jordan, A.S. Willsky, "A Sticky HDP‐HMM with Application to Speaker Diarization," Annals of Applied Statistics, June 2011. (First appeared asMIT LIDS Technical Report P‐2777, November 2007.)

Page 18: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Unsupervised morphological analysis Morphological analysis

To segment sentences into words(morpheme).

This usually requires dictionary(preexisting knowledge of language model).

Unsupervised morphological analysis It does not assume preexisting dictionary.

Mochihashi proposed an unsupervised morphological analysis method based on Nested Pitman‐Yor language model (NPYLM)[Mochihashi ‘09].

Thisisanapple ‐> This is an appleわたしはたなかです. ‐> わたし は たなか です

http://chasen.org/~daiti‐m/paper/nl190segment‐slides.pdf

NPYLM[Moachihashi ‘09]

Page 19: Symbol emergence in robotics @ Shonan meeting 2013/11/13

NPYLM [Mochihashi ‘09]

(Nested Pitman‐Yor Language Model) Mochihashi developed NPYLM for unsupervised morphological analysis.

NPYLM has word n‐gram model and letter ngram model, hierarchically. Each adopts hierarchical Pitman‐Yor language model as a language model (smoothing method).

Bayesian nonparametric model Efficient blocked Gibbs sampler 

Daichi Mochihashi, Takeshi Yamada, Naonori Ueda."Bayesian Unsupervised Word Segmentation with Nested Pitman‐Yor Language Modeling". ACL‐IJCNLP 2009, pp.100‐108, 2009.

Page 20: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Application of Double Articulation Analyzer

DAA

Human motion data

Driving behavior

• Imitation learning [Taniguchi ‘11]• Motion segmentation [Taniguchi’11]

• Extracting driving chunk [Nagasaka ‘12]• Detecting intentional changing points [Takenaka ‘12]• Prediction [Taniguchi ‘12]• Video summarization [Takenaka ‘12]• For topic modeling []

Auditory data*

collaborative workwith DENSO co.

collaborative workwith Nagai lab.

• Language acquisition [Araki ‘12] bottle !!

*Only NPYLM was used.

Page 21: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Contextual Scene Segmentation of Driving Behavior based on Double Articulation Analyzer [Takenaka ‘12] DAA could extract changing points of driving context recognized by human

Kazuhito Takenaka, Takashi Bando, Shogo Nagasaka, Tadahiro Taniguchi, Kentarou Hitomi, Contextual Scene Segmentation of Driving Behavior based on Double Articulation Analyzer, IEEE/RSJ International Conference on Intelligent Robots and Systems 2012 (IROS 2012), 4847‐4852 .(2012)

Page 22: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Semiotic Prediction of Driving Behavior using Unsupervised Double Articulation Analyzer [Taniguchi ‘12]

Tadahiro Taniguchi, Shogo Nagasaka, Kentarou Hitomi, Naiwala P. Chandrasiri, and Takashi BandoSemiotic Prediction of Driving Behavior using Unsupervised Double Articulation Analyzer2012 IEEE Intelligent Vehicles Symposium, 849 ‐ 854 .(2012)

Averaged number of correctlypredicted hidden states

Histogram

Page 23: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Drive Video Summarization based on Double Articulation Structure of Driving Behavior [Takenaka ‘12]

Kazuhito Takenaka, Takashi Bando, Shogo Nagasaka, Tadahiro Taniguchi, "Drive Video Summarization based on Double Articulation Structure of Driving Behavior", ACM multim media 2012,

http://www.youtube.com/watch?v=knwiO6dVbnY

Page 24: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Online Learning of Concepts and Words Using Multimodal LDA and Hierarchical Pitman‐Yor Language Model [Araki ‘12] Connecting multimodal categorization and word segmentation to achieve language acquisition through daily interaction.

Takaya Araki, Tomoaki Nakamura, Takayuki Nagai, Shogo Nagasaka, Tadahiro Taniguchi, Naoto IwahashiOnline Learning of Concepts and Words Using Multimodal LDA and Hierarchical Pitman‐Yor Language ModelIEEE/RSJ International Conference on Intelligent Robots and Systems 2012 (IROS 2012), 1623‐1630 .(2012)

Page 25: Symbol emergence in robotics @ Shonan meeting 2013/11/13

Conclusion Summary

Defining the symbol emergence system Introducing symbol emergence in robotics Introducing double articulation analyzer

Current challenge Unsupervised lexicon acquisition using double articulation analyzer and multi‐modal categorization 

Discussion topic What is the important feature of language which we have to model to obtain computational understanding of human language.


Recommended