Transcript 발표자료
Understanding mirror neurons : A bio-robotic approach 인지과학 협동과정 VCC 석사 2학기 김시온 1.Introduction Understanding action 사과 때리다 왜? 미워서, 웃겨서, 짜증나서 까불어서, 사랑의 마음으로 잘되라고, 말 안들어서… -> 의도 파악의 필요성 1.Introduction Speech understanding machine • Alvin Liberman(May 10, 1917 - Jan. 13, 2000) Analyzing the acoustic characteristics of spoken words Subjects Context Failed because of the limited computational power available Sound Articulatory gestures 1.Introduction Acoustically –evoked motor “Resonance” Production에 관여한 지식들이 Perception에도 관여한다. Action과 관련하여 이러한 역할을 하는 것으로 보이는 영역이 발견되었다!! 2. Physiological properties of monkey rostroventral premotor area (F5) Inferior premotor area 2. Physiological properties of monkey rostroventral premotor area (F5) Motor Neurons • Goal-Directed action들에 반응 (grasping, manipulating, tearing, holding) -비슷한 손동작이 사용될지라도 목적이 다르면 발화하지 않음 • 많은 Grasping neuron들은 특정한 종류 의 Grasping에만 선택적 반응 (precision, finger prehension, whole hand grasping) -> a repository of motor actions 2. Physiological properties of monkey rostroventral premotor area (F5) Visuomotor Neurons - Canonical Neurons • 움직임이 있기 전에, 움직일 수 있는 대상을 보았을 때에도 발화 • 많은 canonical neurons들은 하나, 혹은 소수의 대상에 선택적으로 반응한다. 2. Physiological properties of monkey rostroventral premotor area (F5) Visuomotor Neurons - Mirror Neurons • 내가 행동할 때 뿐만 아니라, 다른 행위자가 동일한 목표-지향적 행동을 할 때에도 발화 • 최근 연구 결과 F5 이외에도 PF에서도 발견됨 • 다른 행위자가 손이나 입을 이용해 물체와 상호작용할 때 활발하게 반응 • 도구를 이용했을 때에는 반응하지 않음 • 대부분의 Mirror Neuron들은 한 종류의 행동에만 선택적으로 반응 (Highly congruent neurons) 2. Physiological properties of monkey rostroventral premotor area (F5) Visuomotor Neurons - Mirror Neurons 3. A model of area F5 and the mirror system A role of F5 Visual appearance of the hand Object affordance Controller Classifier 3. A model of area F5 and the mirror system Controller-predictor formulation Acting Desired action, timing Specific Motor Plan Comparison F5 forward information ≒ Sensory feedback(VMM) Desired action Context ※ Feedback : to counteract disturbances and to learn from mistakes 3. A model of area F5 and the mirror system Controller-predictor formulation Observing an action Transforming visual cues into motor information as before Mirror Neurons Bayesian approach?? 3. A model of area F5 and the mirror system Ontogenesis of mirror neurons • 아기들이 행위자의 목표를 판단할 때 사용하는 Cue는? (Path? Object?) • Results 6,9 mos : new>old goal Woodward(1998) 3. A model of area F5 and the mirror system How do we achieve an action recognizing? Specific goals Canonical Neuron Mirror Neuron 1) Certain actions are more likely to be applied to a particular object 2) Objects are used to perform actions Linking Particular action = particular effect Question!! • 과연 VMM 정보가 Action 인식에 도움을 줄까?? • 로봇를 가지고 한번 직접 실험해 보자! 4. A machine with hands Motor-kinesthetic information + visual information vs only visual information Recording “motor awareness” • Cyber-glove, a pair of CCD cameras, a magnetic tracker, two touch sensors • Data was sampled at frame rate, synchronized and stored to disk by a Pentium class PC • Cyber-glove has 22 sensors , recording the kinematics of the hand at up to 112Hz • Two touch sensors were mounted on the thumb and index finger to detect the moment of contact with the object • Matlab was employed for post-processing 4. A machine with hands Grasping type followed Napier’s taxonomy 4. A machine with hands • Each grasping action was recorded from six different subjects. • Moving the cameras to 12 different locations • Images of the scene from the two camera synchronized with the cyber-glove and magnetic tracker data => Bayesian Classifier • The visual features were extracted from pre-processed image data • Using a simple color segmentation algorithm • The data set was filtered through Principal Component Analysis Classification Task Motor-kinesthetic information + visual information vs only visual information 4. A machine with hands VMM • Using a simple back-propagation neural network with sigmoidal units • The input of the VMM was the vector of the mapping of the images onto the space spanned by the first N PCA vectors • The output was the vector of joint angle acquried the data glove. 4. A machine with hands Experiment 1 - Classification > -> 운동정보가 Action recognition에 도움을 줄 수 있다!!! 5. Robotic experiment Cog Flipper A mirror neuron-like representation could be acquired by simply relying on the information exchanged during the robot- environment interaction (로봇-환경 간의 상호작용에서 mirror-neuron과 같은 표상을 획득할 수 있는가?) 5. Robotic experiment Causal indirection, brain area and function 1 If any neural unit is active in these two situations(both when acting and observing) Then it can be regarded in all respects as a “mirror” unit. by repetitive exposure and manipulation 6. Learning object affordances Rolling affordances Grasping(x) Touching, poking, prodding, sweeping(o) 6. Learning object affordances Rolling affordances 6. Learning object affordances Generic poking actions Can this information be re-used when acting to generate anything useful Showing exploitation of the object affordance? 100/15 … 3번만 완전 실패( ≒ Canonical Neuron) Another interesting question is whether knowledge about object directed actions can be reused in interpreting observed actions performed by a human experimenter. 7. Developing mirror neurons mimicry Same segmentation procedure section 6 generated by human and robot … LEARNING?? 7. Developing mirror neurons Passive vs Active imitation Demiris & Hayes, 2002 7. Developing mirror neurons mimicry Learning of mirror neurons can be carried out by a process of autonomous development 7. Conclusion • The motor information plays a role in the recognition process • A mirror-like representation can be developed autonomously on the basis of the interaction between an individual and the environment 감사합니다.