Successfully reported this slideshow.

Untitled 1

311 views

Published on

  • Be the first to comment

  • Be the first to like this

Untitled 1

  1. 1. Grounding language in action and perception: From cognitive agents to humanoid robots Angelo Cangelosi
  2. 2. Focus <ul><li>Modeling cognitive systems in grounded manner </li></ul>
  3. 3. Grounding of Symbols <ul><li>Symbol learning </li><ul><li>Grounding into categorical representations </li><ul><li>Perceptual categories
  4. 4. Sensorimotor categories </li></ul></ul><li>Symbol relationship </li><ul><li>Combining symbols to form new meanings
  5. 5. Indirect grounding
  6. 6. Addresses symbol grounding problem </li></ul></ul>
  7. 7. Symbol Grounding Problem
  8. 8. Distinction of Model <ul><li>Categorical representations emerge </li><ul><li>Sensorimotor tasks of agents </li><ul><li>Ilkay </li></ul><li>Imitate a teacher </li><ul><li>Nilgun </li></ul></ul></ul>
  9. 9. Distinction of Model <ul><li>ANN based control architecture </li><ul><li>No need to postulate syntactic rules and roles
  10. 10. Able to emerge syntactic structures through learning </li></ul></ul>
  11. 11. Multi-Agent Modeling of Grounding <ul><li>Based on simulated multi-agent system
  12. 12. Population of agent </li><ul><li>Interacts with each other
  13. 13. Sensorimotor and cognitive capabilities through evolutionary learning </li></ul></ul>
  14. 14. Agents <ul><li>Simple organisms </li><ul><li>No realistic representation of body
  15. 15. Perceptual visual input
  16. 16. Motor units to navigate the world and to perform actions </li></ul><li>Perform foraging tasks
  17. 17. Communication with others about interactions </li><ul><li>Location of foods
  18. 18. Performed actions </li></ul><li>Symbolic theft hypothesis describes this evolutionary model </li></ul>
  19. 19. Symbolic Theft Hypothesis <ul><li>Two competing ways of category learning </li><ul><li>Sensorimotor toil
  20. 20. Symbolic theft </li></ul><li>Adaptive benefit of STH is described by Cangelosi and Harnad </li><ul><li>Computational model based on evolutionary foraging task </li></ul></ul>
  21. 21. Evolutionary Foraging Task <ul><li>Agents has to learn two different categories of food </li><ul><li>Mushroom-A: To be eaten
  22. 22. Mushroom-B: To have their location marked
  23. 23. Mushroom-AB: Eaten, marked, and returned to </li></ul><li>Mushroom with irrelevant features (C, D, E) ignored </li></ul>
  24. 24. Evolutionary Foraging Task <ul><li>Bir onceki slaytta bahsedilen gorsel ile aciklanacak </li></ul>
  25. 25. Symbol Grounding Transfer <ul><li>Action and language learning in robotic agents </li><ul><li>Acquiring compositional categories from basic categories
  26. 26. Two agents scenario </li><ul><li>Learner and teacher </li></ul><li>Three stages of learning </li><ul><li>Basic action learning
  27. 27. Naming of learned actions
  28. 28. Higher-order learning by grounding transfer </li></ul></ul></ul>
  29. 29. Symbol Grounding Transfer <ul><li>Ornek gorsel </li></ul>
  30. 30. Language Comprehension in iCub <ul><li>Major innovations </li><ul><li>Realistic model
  31. 31. Object manipulation instructions
  32. 32. Modular architecture </li></ul><li>Integrates various capabilities to respond to linguistic instructions </li><ul><li>Vision processing
  33. 33. Speech processing
  34. 34. Motor processing </li></ul></ul>
  35. 35. Language Comprehension in iCub <ul><li>Ornek </li></ul>

×