Emergence of Compositionality:

Dynamical Systems Accounts from Neuro-Robotics Experiments

May 28,  2013, 4PM

인문신양학술관(4) 3 국제회의실

Abstract:

My group has studied how "compositionality" in our cognitive minds can be

developed as consequences of self-organization in neuro-dynamic systems via

consolidative learning of sensory-motor experiences by conducting synthetic

modeling studies utilizing robotics platforms. In this seminar, I introduce

a basic neuro-dynamic model which is mechanized by two principles, the

predictive coding by the error minimization and multiple timescales dynamics

in the neural activity. The basic model had been implemented in a set of

humanoid robotics experiments including imitation learning of others,

learning of goal-directed compositional object manipulation tasks and

developmental learning of actional concepts via associative learning between

proto-language and actions. These experimental results show (1) how

functional hierarchy can be developed for segmenting continuous

sensory-motor flow into a set of primitive patterns and recombining them,

(2) how actional concepts of combining verbs and objective nouns can be

developed with generalization in the distributed representation of the

neural activity, (3) how free actions can be initiated spontaneously. These

results suggest that the compositionality can be naturally developed in

terms of "organic" structures rather than discrete and algebraic ones as the

results of iterative learning accompanied by dense interactions between the

top-down intention and the bottom-up reality.