the birth of doraemon

27
TAIPEI | SEP. 21-22, 2016 Hao Wu, Cheng Hsin Lee 2016/09/21 THE BIRTH OF DORAEMON

Upload: nvidia-taiwan

Post on 09-Jan-2017

183 views

Category:

Technology


1 download

TRANSCRIPT

Page 1: The Birth of Doraemon

TAIPEI | SEP. 21-22, 2016

Hao Wu, Cheng Hsin Lee 2016/09/21

THE BIRTH OF DORAEMON

Page 2: The Birth of Doraemon

2

AGENDA

What’s a Robot

The birth of DoraemonFunctions and Application scenarios

Difficulties and ChallengesAI2 for Service Robot

Page 3: The Birth of Doraemon

3

ABOUT US

潮汐科技(北京)有限公司

Tide Technology (Beijing) Ltd.

2014.4 2015.7 2016.9

Page 4: The Birth of Doraemon

4

WHAT’S A ROBOT

Page 5: The Birth of Doraemon

5

WHAT’S A ROBOT

Perception Thinking Decision-making Implementation

Robot is made of a complex system

Page 6: The Birth of Doraemon

6

WHY DO DORAEMON

All of us have a dream in childhood,to have one Doraemon.

Page 7: The Birth of Doraemon

7

THE BIRTH OF DORAEMON

R&D for over 1 year,10+ professional teams close cooperation,

more than 200 engineers effort,6 times optimizing on prototypes!

Page 8: The Birth of Doraemon
Page 9: The Birth of Doraemon

9

THE FUNCTION OF DORAEMON

Page 10: The Birth of Doraemon

10

APPLICATION PERSPECTIVE

Page 11: The Birth of Doraemon

11

PRESENTATIONListen to reading in familiar voice

Solution:to acquire user’s Voice Model through deep learning method

Practice:Reading for 2~3 minutes(36 sentences)

Application scenarios:Reading news for elder in Children’s voice

Reading story for child in Mother’s voice

Page 12: The Birth of Doraemon

12

PRESENTATIONEmotion Recognition

To analyze emotion via face recognition.

Page 13: The Birth of Doraemon

13

DIFFICULTIES AND CHALLENGES

Tegra TX1

Page 14: The Birth of Doraemon

AI2 FOR SERVICE ROBOTAI2 = AI Application & Integration

Resources1. Algorithm2. Data3. Calculation capability

Machine Learning1.Unsupervised2.Supervised3.Reinforcement learning

AI2 + Robot1.Environment:Perceive2.People:Understand&Communication3.Autonomous Learning

Page 15: The Birth of Doraemon

1.Visual Perception & Understanding

(Supervised Deep Learning)

2.Strategic Dialogue System

(Deep Reinforcement Learning)

PROJECTS

Page 16: The Birth of Doraemon

1. Training Env.

2. Run time Env.

HARDWARE

Page 17: The Birth of Doraemon

OS – Ubuntu 14.04

CUDA – Nvidia

Performance Libarary - Nvidia

for Visual Perception : Framework –Torch (facebook open source), Neural Talk

For Strategic Dialogue System : Framework – ConvNet, SimpleDS

Dataset : Microsoft COCO & ImageNet

SOFTWARE

Page 18: The Birth of Doraemon

Global Opened Dataset Global Opened Model : VGGNet

PROJECT : DATASET & MODEL

Page 19: The Birth of Doraemon

CNN

object recognition

RNN

language model

PROJECT : ARCHITECTURE

Page 20: The Birth of Doraemon

Smart home environment

PROJECT : TRANSFER OF LEARNING

Page 21: The Birth of Doraemon

PROJECT : FINDING

Page 22: The Birth of Doraemon

PROJECT : APPROACHDeep Reinforcement Learning

Page 23: The Birth of Doraemon

PROJECT

Strategic DialogueDeep Reinforcement

LearningStrategic Dialog System

For compliance

Page 24: The Birth of Doraemon

REFERENCE

● CVPR 2015 Paper● Deep Visual-Semantic Alignments for Generating Image Descriptions Andrej Karpathy, Li Fei-Fei● http://www.cs.toronto.edu/~frossard/post/vgg16/● H. Cuayáhuitl. SimpleDS: A Simple Deep Reinforcement Learning Dialogue System. International

Workshop on Spoken Dialogue Systems (IWSDS), 2016● https://www.cs.utexas.edu/~eladlieb/RLRG.html

Page 25: The Birth of Doraemon

TAIPEI | SEP. 21-22, 2016

Thank you

Page 26: The Birth of Doraemon
Page 27: The Birth of Doraemon