4. Survey
2Karl Friston, Michael Breakspear, and Gustavo Deco. Perception and self-organized instability
Thebrain minimizes the free energy of sensory inputs defined bya generative
model2.
FEP(Free-energy principle)
6. Survey
Free-energy
According to Bayes' theorem, we have…
𝐹 𝑠 = 𝑥 −𝑞 𝑥 ∗ log(𝑝( 𝑠 𝑥)) + 𝐷 𝐾𝐿 𝑞 𝑥 ∥ 𝑝(𝑥) 3 (2)
3吉田正俊 & 田口茂. 自由エネルギー原理と視学的意識.
𝐹 𝑞 = 𝐷 𝐾𝐿 𝑞 𝑥 ∥ 𝑝( 𝑥 𝑠) − log(𝑝(𝑠))3 (1)
To minimize Free-energy, we can change
• the value of q in formula (1) (Perceptual inference)
• the value ofa in formula (2) (Active inference)
7. Survey
Example
3吉田正俊 & 田口茂. 自由エネルギー原理と視学的意識.
It is an image taken bycamera, andwe
are getused to think this is what we
always see. Actually, we only can see
image with high resolution around our
gazedirection.
8. Survey
Example
3吉田正俊 & 田口茂. 自由エネルギー原理と視学的意識.
And this is what we will seeactually.
If wearelooking at +mark, we will see
a vague insect at top right corner.
9. Survey
Example
3吉田正俊 & 田口茂. 自由エネルギー原理と視学的意識.
And if welook at top right corner, we
will know there is a butterfly.
10. Survey
Example
3吉田正俊 & 田口茂. 自由エネルギー原理と視学的意識.
Sensory
Input s
Latent reason x
Generative model 𝑔(𝑥, 𝑠)
X1:Butterfly
X2:Moth
18. • Survey: FEP(Free-energy principle)
• Introduction
• Research progress
• Future Plan
• Discussion and DEMO
Outline
19. Introduction
In spite ofso much technological advancements, communication robothaven’t
becomehuman-equivalent partner.
Smart speaker Communication robot
20. To overcomethis problem, several researcheshave studied on lifelikeness and its
effect to communication.
Motion planning byCPG4 Motion planning byRNN5
4Itsuki Doi, Takashi Ikegami, …, Hiroshi Ishiguro et al. A New Design Principle for an
Autonomous Robot.
5Junsik wang, Jinhyung Kim, …, Jun Tani et al. Dealing with Large-Scale Spatio-Temporal Patterns
in Imitative Interaction between a Robot and a Human by Using the Predictive Coding
Framework
Introduction
21. CPGbased controlframework:
To realize lifelike behavior for android “Alter”4, sensory input to detect unpleasant
situation takes important role to realize lifelikeness in this framework.
CPG
NN
Sensor Net
Introduction
Switcher
4Itsuki Doi, Takashi Ikegami, …, Hiroshi ishiguro et al. A New Design Principle for an Autonomous
Robot.
Values of joints
22. RNNbased controlframework:
In this framework, the robot is controlled to decrease the error between the prediction
and events that actually occurred5.
5Junsik wang, Jinhyung Kim, …, Jun Tani et al. Dealing with Large-Scale Spatio-Temporal Patterns
in Imitative Interaction between a Robot and a Human by Using the Predictive Coding
Framework
Introduction
23. In these researches, CPG orRNN are used for generating diverse motion patterns
and thus the problem can be decomposed into two parts:
• Generation ofdiverse patterns
• CPG, RNN…
• Objective ofthe behavior
• Raising hand when people come close
• Smooth movement
• …
Introduction
24. Introduction
In our research,wedevelop a framework wherediverse motion patterns are
generated by roadmap constructed from recordeddata of human-human
interaction and explore the objective function to realize lifelikeness.
Roadmap
CG Agent
Sensor
Values of joints
Predict
Objective function
Information
User
Reaction
information
30. 2DJoint PosVideo 3DJoint Pos
Research progress
Weused some OSS to extract 3D joint positions from video,
and calculate joint angles for constructing roadmap.
OpenPose6 3D Pose Baseline7
6Zhe Cao, Gines Hidalgo,…, Yaser Sheikh et al. OpenPose: Realtime Multi-Person 2D
Pose Estimation using Part Affinity Fields.
7Julieta Martinez, Rayat Hossain,..., James J. Little. A simple yet effective baseline for
3d human pose estimation
Roadmap
31. Research progress
To construct a roadmap that can be used as motion pattern generator, some rules is
need…
Roadmap
• Initial connection: all states is connected by its
original orderin video.
• Hierarchical Clustering: weremovestates that
are too close to other states.
32. Research progress
To construct a roadmap that can be used as motion pattern generator, some rules is
need…
Roadmap
• Probabilistic connection: newconnection will be
stablished if theirvelocities and accelerations are
similar. And LSH can largely decrease
computational cost in this process.
• Remove isolated states: wewill removestates with
no out-degree recurrently.
33. Research progress
Predict future human action accordingto past information of robotand human
Objective function
ℎ(𝑡 + 1) = 𝑓(ℎ(𝑡), ℎ(𝑡 − 1) … , 𝑟(𝑡), 𝑟(𝑡 − 1) … )
• Candidate methods
• State space
• VAE(Variational Autoencoder)
• Candidate actions tobe predicted
• Gazedirection
• Emotion
• Motion
34. Research progress
State space
I
E
N
S
• State transfer: 𝑥′
= 𝑇𝑥
• Generative model: 𝑝(𝑥, 𝑠) can be calculated from dataset
• Current inference: q( 𝑥 𝑏) can be calculated from T
• Transfer matrix T: 𝑇 = argmin
𝑇
𝐹(p(x, s), q′( 𝑥 𝑏))
• Action: choose action a = argmin
𝑎
𝐹(p(x, s), q′( 𝑥 𝑏) )
State space
36. Research progress
OpenFace
OpenFace
Camera Img
OpenFace is being usedin our research for gaze direction and emotion
detection.
• Gazedirection: can be detected by OpenFace.
• Emotion detection: weconstruct a mapping functionfrom
AUs(Action Units) to emotion9.
9 Sudha Velusamy, Hariprasad Kannan,…, & Bilva Navathe.A method to infer emotions from
facial Action Units
37. • Survey: FEP(Free-energy principle)
• Introduction
• Research progress
• Future Plan
• Discussion and DEMO
Outline
38. Future plan
And this is wewill do next…
• Objective function
• Conversation
• Experiment
39. • Survey: FEP(Free-energy principle)
• Introduction
• Research progress
• Future Plan
• Discussion and DEMO
Outline
40. Discussion and DEMO
This is what we havedone…
• Developed a CG agent system
• Constructed Roadmap as a pattern
generator
• A simple model of Android ibuki
Generative model is a dynamics model produced by past experience of agent. It is usually specified in terms of the likelihood of getting some data given their causes and priors on the parameters.