Page 1
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Bayesian Action-Perception loop modeling: Application to trajectory generation and recognition using internal motor simulationE. Gilet(1), J. Diard(2), R. Palluel-Germain(2), P. Bessière(1)
(1) Laboratoire d’Informatique de Grenoble – CNRS, France(2) Laboratoire de Psychologie et NeuroCognition – CNRS, France
July, 5, 2010http://diard.wordpress.com/ [email protected]
Page 2
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Perception of actions
2
(Calvo-Merino et al., 2004)
Page 3
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Reading and writing letters
3
(Longcamp, 2003)
Writing
Reading pseudo letters
Reading letters
Page 4
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Interpretation
• Motor simulation of actions during perception
• Articulation between perception and action processes
4
Page 5
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Modeling both reading and writingModeling internal simulation of
movements
5
Page 6
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Bayesian Action-Perception (BAP) model
6
Page 7
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Summary
• BAP model – architecture and definition: overview
• Experimental results– simulation of cognitive tasks
• Experimental prediction
7
Page 8
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
BAP model structure
8
internal letter representation
perception model
action model
simulated perception model
coherence variables
Page 9
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Cartesian and effector spaces
• Common space for perceptive and motor internal representations– Cartesian space
9
Page 10
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Letter representation: sequences of via-points
10
Page 11
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
11
Letter representation
« Laplace succession laws »
Page 12
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Parameter indentification
12
Page 13
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
13
Page 14
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
14
Page 15
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
15
Page 16
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
16
Page 17
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Perception model
17
• Deterministic via-point extraction
Page 18
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
18
Action model
Page 19
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
19
Trajectory generation model
• Minimum-acceleration model:– Cost function– Boundary conditions
• Polynomial solution
Page 20
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
20
Simulated perception model
• Identical to the perception model
Page 21
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
21
Coherence variables
• Allow to activate or deactivate submodels– « Bayesian switch »
Page 22
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Coherence variable for controlling submodel activation
• Model– λ binary variable– Joint–
• Inference– P(A) = P(A): value of B does not influence A–
22
A B
λ
A B
A B
Page 23
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Summary
• BAP model – architecture and definition: overview
• Experimental results– simulation of cognitive tasks
• Experimental prediction
23
Page 24
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Perception: reading letters
24
Correct recognition: 93.36%
Page 25
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Perception: writer recognition
25
Correct recognition: 79.5%
Page 26
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Action: writing letters
26
Variability between writers Variability between trials
Page 27
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Motor equivalence
27
Page 28
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Motor equivalence
• Writer “style”– (Wright, 1990)
• Common activated motor areas– (Wing, 2000)
28
(Serratrice. 1993)
Page 29
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Action: Motor equivalence
29
Page 30
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
30
Action: Motor equivalence
Page 31
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Perception and Action: Copy
31
Trajectory copy Letter copy
Page 32
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Perception and Action: Reading letters with motor simulation
32
Recall: reading letters without simulation
Page 33
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
33
Perception and Action: Reading letters with motor simulation
Page 34
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
34
Perception and Action: Reading letters with motor simulation
• Complete trajectories– Correct recognition score with simulation 93.36%– Correct recognition score without simulation 90.2%
• Incomplete trajectories
Page 35
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Summary
• BAP model – architecture and definition: overview
• Experimental results– simulation of cognitive tasks
• Experimental prediction
35
Page 36
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Experimental prediction
36
Page 37
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Preliminary data
37
60
70
80
90
Control Group (motor simulation unaffected)
Motor interference Group (motor simulation affected)
Complete letters
Truncated letters
Recognition Performance (%)
F(1,23) = 3.06, p = 0.093
Page 38
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Summary
• BAP model– Bayesian model of perception
and action– Includes an internal
simulation loop• Cognitive tasks
– Reading without and with motor simulation
– Writer recognition– Writing with different
effectors– Copying letters and
trajectories• Basis for experimental
predictions38
Page 39
Gilet, Diard, Palluel-Germain & Bessière — LIG & LPNC-CNRSBayesian Action-Perception model
Thank you for your attention !
Questions ?