HAL Id: hal-01963115https://hal.inria.fr/hal-01963115

Submitted on 21 Dec 2018

HAL is a multi-disciplinary open accessarchive for the deposit and dissemination of sci-entific research documents, whether they are pub-lished or not. The documents may come fromteaching and research institutions in France orabroad, or from public or private research centers.

L’archive ouverte pluridisciplinaire HAL, estdestinée au dépôt et à la diffusion de documentsscientifiques de niveau recherche, publiés ou non,émanant des établissements d’enseignement et derecherche français ou étrangers, des laboratoirespublics ou privés.

Learning an inverse model for vocal production: towarda bio-inspired model

Silvia Pagliarini, Xavier Hinaut, Arthur Leblois

To cite this version:Silvia Pagliarini, Xavier Hinaut, Arthur Leblois. Learning an inverse model for vocal production: to-ward a bio-inspired model. European Birdsong Meeting, Apr 2018, Odense, Denmark. �hal-01963115�

Learning an inverse model for vocal production: toward a bio-inspired model

6th European Birdsong Meeting, April 12-13, 2018, Odense, Denmark

Silvia Pagliarini (with Xavier Hinaut and Arthur Leblois)INRIA Bordeaux Sud-Ouest, Institut des Maladies Neurodégénératives, Université de Bordeaux, FR

WHAT IS SENSORIMOTOR LEARNING?

Control problem which maps a sensory input into a motor output

Basic components:

● Input: sensory stimulus

● Output: reproduction of the stimulus

Da Cunha et al., 2010

LEARNING BY IMITATION AND INVERSE MODEL

Imitation: learning from a tutor using a feedback guided error

Sensory area Motor area

Motor production

LEARNING BY IMITATION AND INVERSE MODEL

Imitation: learning from a tutor using a feedback guided error

Inverse model: the aim is to transform a sensory stimulus into the corresponding motor command

Sensory area Motor area

Inverse model

Motor production

A BIOLOGICAL EXAMPLE: SONG LEARNING IN BIRDS

Sensory Subsong (Babbling) Plastic song Crystallization

● Comparable learning mechanisms and behavior

Brainard and Doupe, 2002

A BIOLOGICAL EXAMPLE: SONG LEARNING IN BIRDS

● Comparable learning mechanisms and behavior

Sensory Subsong (Babbling) Plastic song Crystallization

Brainard and Doupe, 2002

A BIOLOGICAL EXAMPLE: SONG LEARNING IN BIRDS

● Comparable learning mechanisms and behavior

Sensory Subsong (Babbling) Plastic song Crystallization

Brainard and Doupe, 2002

LEARNING AN INVERSE MODEL

Synaptic weights initially weak

Sensory Area

Motor Area

…

…

LEARNING AN INVERSE MODEL

Synaptic weights initially weak

At each time :

●

Sensory Area

Motor Area

…

…

LEARNING AN INVERSE MODEL

Synaptic weights initially weak

At each time :

●

●

: learning rate

Sensory Area

Motor Area

…

…

Hebbian learning rule

LEARNING AN INVERSE MODEL

Synaptic weights initially weak

At each time :

●

●

●

: learning rate

Sensory Area

Motor Area

…

…

Hebbian learning rule

HAHNLOSER-GANGULI THEORETICAL MODEL

Time (in number of time steps)

Ave

rage

dis

tanc

e ov

er 5

0 s

imul

atio

ns

NONLINEAR MODEL INTRODUCTION

: target motor pattern

: tuning selectivity width

represents the distance between the target and the random exploration

GANGULI-HAHNLOSER MODEL

Time (in number of time steps)

Ave

rage

dis

tanc

e ov

er 5

0 s

imul

atio

ns

NORMALIZATION

Synaptic weights have a maximal value, related to the number of synaptic receptors one neuron is able to produce.

NORMALIZATION

Synaptic weights have a maximal value, related to the number of synaptic receptors one neuron is able to produce.

● Maximal weights normalization

● Supremum weights normalization

NORMALIZATION

Synaptic weights have a maximal value, related to the number of synaptic receptors one neuron is able to produce.

● Maximal weights normalization

● Supremum weights normalization

● Decreasing factor normalization

NORMALIZED INVERSE MODEL

Normalization applied over the auditory neurons

Maximum weights

Supremum weights

Decreasing factor

Time (in number of time steps)

Evol

utio

n of

the

dist

ance

NORMALIZED INVERSE MODEL

Norm

Time (in number of time steps)

Evol

utio

n of

the

dist

ance

Mean

AUDITORY SELECTIVITY EFFECTC

onve

rgen

ce ti

me

(in n

umbe

r of t

ime

step

s)

Tuning selectivity width

Dis

tanc

e fr

om th

e ta

rget

VARYING INPUT/OUTPUT DIMENSION

Distance from the motor target at convergence

Dis

tanc

e fr

om th

e ta

rget

Number of neurons in the network

VARYING INPUT/OUTPUT DIMENSION

Convergence time

Con

verg

ence

tim

e (in

num

ber o

f tim

e st

eps)

Number of neurons in the network

SUMMARY

● Simple normalization schema are successful in the nonlinear model

● Decreasing tuning selectivity width: ○ convergence time explosion ○ accuracy of learning increases

● Auditory VS motor dimension

WHAT’S NEXT?

● Duration of syllable and feedback delay

WHAT’S NEXT?

● Duration of syllable and feedback delay

● Production of sound

WHAT’S NEXT?

● Duration of syllable and feedback delay

● Production of sound Enjoy the poster from Xavier Hinaut

WHAT’S NEXT?

● Duration of syllable and feedback delay

● Production of sound

● Make prediction on experimental data

Enjoy the poster from Xavier Hinaut

Thanks for the attention.