A bio-inspired model towards vocal gesture learningin songbird

Silvia Pagliarini1,2,3, Xavier Hinaut1,2,3, Arthur Leblois3

1) Mnemosyne, Inria Bordeaux Sud-Ouest2) LaBRI, UMR 5800, CNRS

3) IMN,UMR 5293, CNRSUniversité de Bordeaux, France.

WHAT’S NEXT?

NORMALIZED INVERSE MODELINTRODUCTION

Sensorimotor learning: control problem which maps a sensory input into a motor output.

RESULTS

Imitation: learning from a tutor using a feedback guided error.

Da Cunha et al., 2010

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

VARYING INPUT/OUTPUT DIMENSION

Dis

tanc

e fr

om th

e ta

rget

Number of neurons in the network Number of neurons in the network

Conv

erge

nce

time

(in n

umbe

r of t

ime

step

s)Learning accuracy Learning speed

Motor dimension influences learning making it slower and slower as it increases.

LEARNING AN INVERSE MODEL

SYNAPSIS WEIGHTS LIMITATION

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

LINEAR (R. H. Hahnloser & S. Ganguli [4] ) vs NONLINEAR MODEL

● Exploration of motor dimension, dividing it in term of behavioral output and motor commands.● Implement a realistic vocal production system (motor control model).

Time (in number of time steps)

Dis

tanc

e fr

om th

e ta

rget

● Maximal weights normalization:

● Supremum weights normalization:

● Decreasing factor normalization:

normmean

AIM: implementation of an inverse model which describes the sensorimotor phase of learning in birdsong.

● Sensorimotor phase is characterized by babbling.

● Dedicated vocal circuit:

Learning by imitation

Brainard and Doupe, 2002

Motor pathwayAuditory pathwayLearning pathway

a. Adult zebra finch song (TUTOR).b. Juvenile zebra finch song at an early stage of learning.c. Song close to crystallization.

SONG LEARNING IN BIRDS

KEYWORDS

SUMMARY● Normalization gives better performance when applied over auditory neurons.● Low selectivity tuning width makes learning slower but more accurate.● Motor dimension influences learning making it exponentially increasing.

BIBLIOGRAPHY

DISTANCE FROM THE TARGET AND CONVERGENCE

Distance at each time step: Convergence is reached when the distance reaches a plateau.

(1) Brainard, M. S. and Doupe, A. J., What songbirds teach us about learning, Nature,2002(2) Mooney R., Neural mechanisms for learned birdsong, Cold Spring Harbor Laboratory Press, 2009

(3) Wolpert D. M., Diedrichsen J. and Flanagan J. R., Principles of sensorimotor learning, Nature Reviews, 2011(4) Hahnloser R. H. R. and Ganguli S., Vocal learning with inverse models, Principles of Neural Coding, 2013, CRC Press Boca Raton

Number of motor neurons

Mean of synaptic weights columns

Ideal auditory activity

NORMALIZED INVERSE MODEL

AUDITORY SELECTIVITY ANALYSIS

Selectivity tuning width

Conv

erge

nce

time

(in n

umbe

r of t

ime

step

s)

Dis

tanc

e fr

om th

e ta

rget

High selectivity - which means low tuning width -

makes learning slower but more accurate.

For simulations the decreasing factor

normalization has been applied.

● Linear auditory response:

● Nonlinear auditory response:

Two populations: auditory and motor neurons

selectivity tuning width which drives auditory

selectivity.

At each time step :

● Auditory feedback: ● ● Update weights:

Hebbian learning rule

: learning rate

Motor random exploration

Target motor pattern

Dis

tanc

e fr

om th

e ta

rget

Time (in number of time steps)

Nonlinear model does not converge applying this learning rule.

Dis

tanc

e fr

om th

e ta

rget

Sensory area

Motor area

- Synaptic weights

Selective sensory responseSensory response

Three normalizations

meannorm

Normalization applied over auditory neurons gives better performance.

Auditory neurons

Time (in number of time steps) Time (in number of time steps)

Evol

utio

n of

te w

eigh

ts

Motor neurons

Motor VS AuditoryEvolution of weights

Decreasing factor normalization gives better performance.

Normalization applied over auditory neurons