Michale Fee

McGovern Institute for Brain Research

Department of Brain and Cognitive Sciences

MIT

Jerusalem in Motion Workshop

Jerusalem, Israel

December 18, 2003

Vocal control in the songbird: Neural mechanisms of sequence

generation

A

B

C

D

EF

G

H

I

J

A-B-C-D-E-F-G-H-I-J

1 2 3

Abeles, Hertz, ‘80s and ‘90s

Synchronous Firing Chain

Neural Circuits for Sequence Generation

1 2 3 4

1

23

4

Metastable AttractorsSompolinsky, Kleinfeld, Platt, 1980s

fast

slow

Neural Circuits for Sequence Generation

• Train a specified sequence of neural states,

• Sequence of states must be nearly orthogonal• A-B-C-A-D is not allowed

• Interference between sequence and dynamics

• Timescale is set by synaptic/biophysical time constants

Wij = SitSj

t+1

Sit

t

Overview

• Songbird as a model system

• Technological challenges

• Mechanisms of sequence generation in the songbird

Zebra Finches

0 kHz

10 kHz

Zebra Finch Song Structure

1s

Fre

quen

cy

Motif Motif Motif

Syllable

Songbird Vocalizations are Highly Stereotyped

Songbirds Can Generate Output Over a Wide Range of Timescales

Biological systems can:

• Learn and reliably generate low-dimensional sequential behavior– not a specified sequence of neural states

• Generate an arbitrary sequence– not constrained by orthogonality between output

states

• Operate over a wide range of timescales

Circuits for Vocal Production and Learning

H V C

R A

U VA

XDLM

LM AN

nX IIts

S yrinx

N If

Motor Circuit

Learning Circuit

(7)

1000

7000

20,000

Technical Difficulties

• Songbirds will only sing while unconstrained

• Zebra finch weighs only 12-15 grams

• Singing is suppressed by handling

• 3 independently controlled electrodes

• Motorized for remote control

• 1.5 gram total weight

Motorized Miniature Microdrive

Fee and Leonardo, 2000

Premotor Activity During Singing

Bou

tM

otif

Instantaneous Firing Rate

0.0 0.4 0.6 0.80.2

1

6

12N

euro

n #

Time [s]

Firing R

ate [1 kHz/D

iv]

How Are the Burst Sequences in RA Generated?

• Internal dynamics within RA?

- OR -

• Imposed from HVC?

Models of Pattern Generation in HVC and RA

Fee

d-fo

rwar

dIn

trin

sic

HVC

RA

HVC

RA

~10ms

~10ms

Singing Related Firing Patterns in Nucleus HVC

Yu and Margoliash, 1996

Antidromic Identification of HVC Neurons

X

Stim

StimHV c

RA

What do RA-Projecting HVC neurons do during singing?

Hahnloser, Kozhevnikov, and Fee, Nature (2002)

Hahnloser, Kozhevnikov and Fee, Nature (2002)

Simple Sequence Generation Circuit

Sparse representation of time

Fixed synaptic weights

Plastic synaptic weights

Downstream effect of RA activity

Simple Sequence Generation Circuit

Sparse representation of time

Fixed synaptic weights

Plastic synaptic weights

Model of Vocal LearningH

VC

100

110

120

Initi

al o

utpu

tF

inal

ou

tpu

t

0 50 100 150Tim e (m s)

with Sebastian Seung and Ila Fiete

A Sparse Representation in HVC Speeds Learning

0 5 10 15 20 25 30

10-2

10-1

100

Sq

uare

d e

rror

Learn ing iterations

1

248

with Sebastian Seung and Ila Fiete

Simple Sequence Generation Circuit:Emergent RA activity

Emergent Activity in RA Neurons

with Sebastian Seung and Ila Fiete

Emergent Activity in RA Neurons

• Each model RA neuron has a unique pattern of bursts

• A different ensemble of active RA neurons at each time in the sequence

• The ensemble of active RA neurons evolves to an uncorrelated ensemble every ~10 ms, even during constant output

RA ensembles are uniquely related to a temporal position in the output

– not to motor output

How is this possible?

High Degree of Convergence From RA to Motor Output

• ~7000 RA projection neurons

• ~1000 motor neurons

• 7 muscles

Many Different Ensembles of Active RA Neurons Can Produce the Same Motor Output

Model RA outputs form a highly degenerate code for motor signals

RA

Instantaneous Firing Rate

0.0 0.4 0.6 0.80.2

1

6

12N

euro

n #

Time [s]

Firing R

ate [1 kHz/D

iv]

Tim

e t 2

020

040

060

01 25

Neuron #

Time t1

0 200 400 600

1

25

Neu

ron

#

Time t1

Tim

e t 2

0 200 400 600

020

040

060

0

How are the Timescales of Neural and Motor Activity Related?

Neural and Song Correlation Matrices

Neural and Song Correlation Width

Circuits for Vocal Production and Learning

H V C

R A

U VA

XDLM

LM AN

nX IIts

S yrinx

N If

Motor Circuit

Learning Circuit

• Each RA neuron has a unique pattern of bursts

• A different ensemble of active RA neurons at each time in the song motif

• The ensemble of active RA neurons evolves to an uncorrelated ensemble every ~10 ms, even during parts of the song with constant acoustic output

Our proposed network can:

• Learn and reliably generate low-dimensional sequential behavior– not a specified sequence of neural states

• Generate an arbitrary sequence– not constrained by orthogonality between output

states

• Operate over a wide range of timescales

Design Principles and Implications

• Separate the temporal dynamics and the mapping to motor output – Changes in learned output do not affect temporal

structure

• Sparse coding of temporal order in HVC– Fast learning?

– No single neuron tuning in RA?

Future Directions• When during development does the sparse

representation of time in HVC arise?

• Where do sparse sequences in HVC originate? Intrinsic dynamics within HVC, or driven from NIf?

Circuits for Vocal Production and Learning

H V C

R A

U VA

XDLM

LM AN

nX IIts

S yrinx

N If

Motor Circuit

Learning Circuit

1 2 3

Where and how is ‘time’ generated?

1 2 3 4

1

23

4

fast

slow

Collaborators

• Richard Hahnloser– Bell Laboratories

• Alexay Kozhevnikov– Bell Laboratories

• Anthony Leonardo– Bell Laboratories

• Ila Fiete, Sebastian Seung

– Brain and cognitive sciences department – MIT

Simple Models of Neural Circuits

1 2

1

21

1

A

B

• stable states - fast, symmetric connections

1

2

1 2

• dynamic states - slow or asymmetric connections