Diese Präsentation wurde erfolgreich gemeldet.
Wir verwenden Ihre LinkedIn Profilangaben und Informationen zu Ihren Aktivitäten, um Anzeigen zu personalisieren und Ihnen relevantere Inhalte anzuzeigen. Sie können Ihre Anzeigeneinstellungen jederzeit ändern.

Wei Xu at AI Frontiers : Language Learning in an Interactive and Embodied Setting

352 Aufrufe

Veröffentlicht am

Language Learning in an Interactive and Embodied Setting

Veröffentlicht in: Technologie
  • Als Erste(r) kommentieren

  • Gehören Sie zu den Ersten, denen das gefällt!

Wei Xu at AI Frontiers : Language Learning in an Interactive and Embodied Setting

  1. 1. Horizon Robotics Language Learning in an Interactive and Embodied Setting 11/2018 Wei Xu 1 Horizon Robotics
  2. 2. Horizon RoboticsA Developmental Approach to Machine Intelligence 1. It might be easier than solving all the tasks a human adult can do 2. Learn skills and knowledges unspecified at design time 3. Gradually proceed from easy tasks to difficult tasks 2 “Instead of trying to produce a program to simulate the adult mind, why not rather try to produce one which simulates the child's? If this were then subjected to an appropriate course of education one would obtain the adult brain.” - Alan Turing (1950) Language learning in an interactive and embodied setting
  3. 3. Horizon RoboticsWhy Embodied?  Learn from the experiences coming from the machine’s interactions with its environment  Learn commonsense through the observation and interaction with the environment  Meaning emerges by “grounding” language in modalities in our environment 3Language learning in an interactive and embodied setting Human driving: < 1000 miles Self-driving: >10 million miles
  4. 4. Horizon RoboticsWhy Interactive?  A useful robot needs to be able to understand and communicate effectively  It is easier for human to teach machines directly using language than writing code  Humans are great teachers  Learn the effects of speaking by observing feedbacks from conversational partner  Learn human value through the interaction 4Language learning in an interactive and embodied setting
  5. 5. Horizon RoboticsAnswering Questions and Following Commands 1. Is it possible to learn to follow commands using end-to-end reinforcement learning without any pretraining for vision or language? 2. Whether learning question answering can help learning command 3. Can the machine understand words under new context not seen in training? 5 Haonan Yu, Haichao. Zhang, Wei Xu “Interactive Grounded Language Acquisition and Generalization in a 2D World” ICLR 2018
  6. 6. Horizon RoboticsProblem Setup 6Answering questions and following commands east and avocado never appears together in training Watermelon only appears in answers during training
  7. 7. Horizon RoboticsModel architecture 7Answering questions and following commands answer action value
  8. 8. Horizon RoboticsExperiments 8Answering questions and following commands No QA training
  9. 9. Horizon RoboticsGeneralization Ability 9 We can generalize to word combinations never seen in training We can generalize to questions containing words never seen in training Answering questions and following commands Held out X(%): %X of word/combinations are held out from training
  10. 10. Horizon Robotics Challenges:  Partially observed  Much longer delay of reward  More visual variations “Navigate to the dog!”Navigation in a 3D Environment 10
  11. 11. Horizon RoboticsGuided Feature Transformation Haonan Yu, Xiaochen Lian, Haichao Zhang, Wei. Xu “Guided Feature Transformation (GFT): A Neural Language Grounding Module for Embodied Agents” CoRL 2018 11Navigation in 3D environment action value
  12. 12. Horizon RoboticsExperimental Results 12Navigation in 3D environment
  13. 13. Horizon RoboticsDemo the object besides candle is your target . please move to the object that is front of the basketball . can you reach the object right of toilet ?go to the object to the right of bike please .reach the location between car and trampoline please.please navigate to the grid between gift and tower .please navigate to the grid between bucket and chair .please move to the object that is front of basketball . 13Navigation in 3D environment
  14. 14. Horizon RoboticsLearning to Speak and Remember 1. How to learn to speak by talking with other people? 2. What information should be remembered? 3. How to utilize knowledge in memory? 14 Haichao Zhang, Haonan Yu, Wei Xu “Interactive Language Acquisition with One-Shot Visual Concept Learning through a Conversation Game” ACL 2018
  15. 15. Horizon RoboticsProblem Setup Rewards are given for each learner response based on its appropriateness 15Learning to speak and remember
  16. 16. Horizon RoboticsMemory Augmented Imitation + Behavior Shaping Through RL Interpreter Speaker Vision Memory What is this? It is a bird. Reward 16Learning to speak and remember
  17. 17. Horizon RoboticsModel Detail 17 Trained end-to-end using gradient descent over Imitation Cost + Reinforce Cost Learning to speak and remember
  18. 18. Horizon RoboticsExample Dialogs 18Learning to speak and remember T: Virtual teacher L: Learner (machine) T: i see grape L: watermelon grape watermelon T: tell what you see L: see see see see see T: there is grape L: grape grape watermelon T: i can observe coconut L: fox watermelon watermelon ------------------------------------------------- _________________________________ ------------------------------------------------- ------------------------------------------------- _________________________________ Before learning After learning
  19. 19. Horizon RoboticsSummary  What we have now:  Learning to understand and use simple language, memorize useful information, and execute simple commands from the interactions with a virtual teacher in virtual environments  What we will do in the future:  Simple → complex  Virtual → real 19
  20. 20. Horizon RoboticsAI Research at Horizon Robotics  About the company  A leading technology powerhouse of edge AI platform  Provide algorithms, processors and hardware jointly optimized for high-performance, low- power and low-cost edge AI capabilities  CES 2019 Innovation Reward  General AI Lab @ Silicon Valley  Research towards the company’s long term vision for artificial general intelligence  Build machines that can learn skills and knowledges unspecified at design time  Applied AI Lab @ Silicon Valley  Applied research focusing on near term needs  Developing novel AI technologies that are critical to our current products Job: bit.ly/general-ai-lab bit.ly/applied-ai-lab 20
  21. 21. Horizon Robotics THANKS! 21