soinn : an artificial brainwatanabe-...3 1. understand a voice command 2. learn autonomously from...

54
SOINN : An Artificial Brain Osamu Hasegawa

Upload: others

Post on 18-Mar-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN : An Artificial Brain

Osamu Hasegawa

Page 2: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

2

Motivation

hard-cording

operator

Confinedenvironment

Page 3: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

3

1. Understand a voice command

2. Learn autonomously from the Internet, if needed

3. Generate the procedure and accomplish the task.

We need to develop a robot which …

Page 4: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Autonomous Mental DevelopmentWeng, J., et. al. (2001). Autonomous mental development by robots and animals., Science, 291, 599-600.

Page 5: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Big Data• “Infoplosion” is underway globally on daily basis. Data

accumulation volume is enormous.

• Computers and robots in addition to human beings should use these data for learning, associating, reasoning, knowledge-transferring and forecasting.

• We want to eliminate irrelevant noise automatically.

• SOINN does it!

Page 6: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

6

Major Advantages of SOINN :

1. SOINN user does not need to define a mathematical model for learning.

2. SOINN can automatically eliminate noise.

Page 7: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

7

Self-organizing Incremental Neural Network (SOINN)

Noise is eliminated. Pattern is recognized.

Page 8: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN applet demo

8

Page 9: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN Algorithm

9

Page 10: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN has vast possibilities

• SOINN is algorithm: it works with any language: C, C++, C#, JAVA, GP-GPU, Matlab

• SOINN works with any hardware:– Google Glass, Cellular phone, smart phone, personal

computer, embedded system, etc.

• SOINN works best with internet:– Google Drive, finance, distribution,

climate/environment, education, medicine, etc.

10

Page 11: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOIAM• SOIAM: SOINN for Associative Memory• Learn a pair of vectors (a key and an

associative vector)• Realize the association between the

above pair• One-to-many and many-to-many

association

11Example of one-to-many association

Page 12: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOIAM (conceptual diagram)

12

associative vector

key vector

Node of SOINN

Page 13: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN-PBR

• SOINN-PBR: SOINN for Pattern-Based Reasoning

• Realize reasoning with the pattern-based if-then rules of propositional logic– Furthermore, realize deductive inference by

using the known pattern-based if-then rules• Extend SOIAM to learn the pattern-based

if-then rules

13

Page 14: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

A∧B→CA→B∨C

SOINN-PBR(conceptual diagram)

14

conditional rule

sequential rule

Edge for ‘AND’

Page 15: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN for Real-world Robots

15

Page 16: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

UNGUIDED ROBOT NAVIGATION USING CONTINUOUS ACTION SPACESirinart Tangruamsub, Manabu Tsuboyama, Aram Kawewong and Osamu Hasegawa

Page 17: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

If-then Rules• Use for searching path fragments (from start to goal).

• where• q = ‘from-position’ or ‘current position’• a = ‘robot action’• f = ‘to-position’

q ^ a f

Do an action aq

fqa

Page 18: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN Architecture

Page 19: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN Architecture• Input Layer• Role:

• Receive data from robot• We separated this layer

into 2 parts• Image data (q,f) • Robot action data

• Memory Layer• Role:

• Memorize the data pattern from the input layer

Page 20: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN Architecture• Associative Layer• Role:

• Shows the relation ship between image data and action data (if-then rules).

^q ^ a f

Page 21: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Experiment 1• Robots automatically learns the associations between images including actions

Page 22: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Experiment 1• Autonomous path planning

想起 推論1 推論2

推論3

推論4推論5推論6

入力データ内部データ

#6253内部データ

#6303内部データ#11719

内部データ#16564

内部データ#8193

内部データ#58194

内部データ#58278

左:1.0右:-1.0(右旋回)

左:1.0右:-1.0(右旋回)

左:1.0右:0.6(斜め右)

左:1.0右:1.0(前進)

左:1.0右:0.6(斜め右)

左:1.0右:-1.0(右旋回)

Start

Goal

Page 23: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Experiment 2

Page 24: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Experiment 2• Training

+“さん”

+“よん”

+“いち”

+“に”

Page 25: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Experiment 2

Autonomous path planning from 「いち」 to 「に」

Page 26: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

26

Experiment in crowded environment

Page 27: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need
Page 28: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Autonomous Robot Navigation

Morioka 2010

Page 29: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Proposedw/o Prop.Method

Final results

Page 30: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Mate Tea Cup …??

30

Page 31: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN + internet• With internet SOINN learns instantly

and becomes wiser.– High-speed learning and online learning

• In the ever-changing real world environment, it is extremely difficult to preliminary assume effective attributes.

• SOINN extracts and decides effective attributes by interacting with human beings and the environment.

– Internet information is extremely noisy.• SOINN is highly noise-resistant.

Page 32: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Transfer Learning • SOINN learns basic concepts (“attributes”),

not objectives• SOINN recognizes non-pre-learned

objectives by combining attributes.

boxes

Box-style thing with dial and window= microwave

balls

window

microwaves

Defined by human beings

Images for learningdials

transfer

Page 33: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

33

Alphabet and Dictionary.

Page 34: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Tomato Broccoli

Banana Cinnamon

Image examples for experiments

Page 35: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

35

“Guess unknown objects”

Page 36: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

36

Results

CVPR09[1]

IJCNN11[2]

CVPR09[1]

IJCNN11[2]

70 days

6 hours

7 mins

Test time(Approx. 6000imgs) 2 days

4 hours

90 secs

Training time(Approx. 25000imgs)

Without time to extract features

Proposed

Proposed

Page 37: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

37

Page 38: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

38

We have many modalities.

Page 39: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

39

Page 40: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

40

Modalities of SOINN Robot

1. Image2. Sound3. 3D-Info4. Pressure5. Weight

Page 41: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Pressure

Sound

Image

Depth

Weight

Mate tea cup

SOINN robot also learn Weight of attributes.

Wooden-attr

ImageDepthSoundPressureWeight

e.g.,

Attribute of “Wooden” is mostly understood

by Sound

Page 42: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Knowledge transfer between robots

Robots teaching robots

Hi, I will tell you how to

make British tea.

Transfer

42

Page 43: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

In the future, people will grow up with her/his own Artificial Brain to supplement brain functions. 43

Page 44: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

MY Artificial Brain,Always with Me !

44

Page 45: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

Appendix

45

Page 46: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINNによる超高速・汎用確率密度推定

𝒩𝒩 :ノードの全体集合

𝒘𝒘𝑛𝑛:ノード 𝑛𝑛の位置ベクトル

𝑡𝑡𝑛𝑛 :ノード 𝑛𝑛が競合学習で勝者になった回数

𝑇𝑇𝒩𝒩 = ∑𝑛𝑛∈𝒩𝒩 𝑡𝑡𝑛𝑛

𝐾𝐾𝑯𝑯 𝒙𝒙 = 12𝜋𝜋 𝑑𝑑 𝑯𝑯

𝑒𝑒𝑒𝑒𝑒𝑒 − 12𝒙𝒙𝑇𝑇𝑯𝑯−1𝒙𝒙 : Gauss Kernel

�̂�𝑒 𝒙𝒙 = �𝑛𝑛∈𝒩𝒩

𝑡𝑡𝑛𝑛𝑇𝑇𝒩𝒩

𝐾𝐾𝑪𝑪𝑛𝑛 𝒙𝒙 − 𝒘𝒘𝑛𝑛

46

Page 47: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN vs Infinite Gaussian mixture model (IGMM)

Page 48: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN vs Infinite Gaussian mixture model (IGMM)

Page 49: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN vs Infinite Gaussian mixture model (IGMM)

Page 50: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN vs Infinite Gaussian mixture model (IGMM)

Page 51: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

SOINN vs Infinite Gaussian mixture model (IGMM)

Page 52: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need
Page 53: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

実験 : 計算時間

53

S-KDE : 𝐻𝐻 = 𝜎𝜎2𝐼𝐼 (行列計算なし)U-KDE : 𝐻𝐻 (行列計算あり)

0

50

100

150

200

250

300

350

400

10000 20000 30000 40000 50000 60000 70000 80000 90000 100000

計算

時間

(s)

サンプル数

KDE-SOINN

S-KDE

U-KDE

KDE-SOINN : Local Kernel

Page 54: SOINN : An Artificial Brainwatanabe-...3 1. Understand a voice command 2. Learn autonomously from the Internet, if needed 3. Generate the procedure and accomplish the task. We need

0

100

200

300

400

500

600

700

800

900

1000

10000 20000 30000 40000 50000 60000 70000 80000 90000 100000

計算

時間

(ms)

サンプル数

KDE-SOINN

Proposed

実験 : 計算時間(前頁の拡大図)

54

サンプル数が10倍に増えても計算時間は2倍程度の増加に収まっている!

Local Kernel