gain control in insect olfaction for efficient odor recognition
DESCRIPTION
Gain control in insect olfaction for efficient odor recognition. Ram ón Huerta Institute for Nonlinear Science UCSD. The goal. What is time and dynamics buying us for pattern recognition purposes?. One way to tackle it. - PowerPoint PPT PresentationTRANSCRIPT
Gain control in insect olfaction for efficient odor recognition
Ramón Huerta
Institute for Nonlinear Science
UCSD
What is time and dynamics buying us for pattern recognition purposes?
One way to tackle it
1. Start from the basics of pattern recognition: organization, connectivity, etc..
2. See when dynamics (time) is required.
The goal
How does an engineer address a pattern recognition problem?
1. Feature extraction. For example: edges, shapes, textures, etc…
2. Machine learning. For example: ANN, RBF, SVM, Fisher, etc..
What is easy ? What is difficult?
1. Feature extraction: very difficult (cooking phase)
2. Machine learning: very easy (automatic phase)
Feature Extraction
High divergence-convergence ratiosfrom layer to layer.
Antennal Lobe (AL)Mushroom body (MB)
Antenna
Mushroom body lobes
Location of learning
How insects appear to do it
Machine Learning
Stage
Bad news
The feature extraction stage is mostly genetically prewired
Good news
The machine learning section seems to be “plastic”
Feature Extraction
Antennal Lobe (AL) Mushroom body (MB)
Antenna
Mushroom body lobes
Machine Learning
Stage
Spatio-temporal coding occurs here No evidence of time here
The basic question
Can we implement a learning machine with
• fan-in, fan-out connectivities,
• the proportion of neurons,
• local synaptic plasticity,
• and inhibition?
Huerta et al, Neural Computation 16(8) 1601-1640 (2004)
Marr, D. (1969). A theory of cerebellar cortex. J. Physiol., 202:437-
470.
Marr, D. (1970). A theory for cerebral neocortex. Proceedings of the
Royal Society of London B, 176:161-234.
Marr, D. (1971). Simple memory: a theory for archicortex. Phil.
Trans. Royal Soc. London, 262:23-81.
Willshaw D, Buneman O P, & Longuet-Higgins, HC (1969) Non-holographic associative memory, Nature 222:960
CALYX
Display Layer
IntrinsicKenyon Cells
PNs (~800) iKC(~50000) eKC(100?)
AL
MB lobes
Decision layer
ExtrinsicKenyon Cells
No learningrequired
Learningrequired
k-winner-take-all
Stage I: Transformation into a large displayStage II: Learning “perception” of odors
1y
0y
2y
3y
KCs coordinates
0x10
1x00
2x10 3x
11
Class 1 Class 2
1
0 1
0x
2x
3x
1x
AL coordinates
1y
0y
2y
3yw
Hyperplane:Connections from the
KCs to MB lobes
MB lobe neuron:decision
Odor classification
Odor 4
Odor 3
Odor 2
Odor 1
Odor N
Class 1
Class 2
Sparse code P
rob
abil
ity
of d
iscr
imin
atio
n
# of active KCs
Capacity for discriminating
We look for maximum number of odors that can be discriminated for different activate KCs,
Note: we use Drosophila numbers
KCn
# of active KCs
TO
TA
L #
OF
O
DO
RS
It has been shown both inLocust (Laurent)
and Honeybee (Menzel)
the existence of sparse code~1% activity
Narrow areas of sparse activity
Without GAIN CONTROL
There can be major FAILURE
Feature Extraction
Antennal Lobe (AL) Mushroom body (MB)
Antenna
Mushroom body lobes
Machine Learning
Stage
GAIN CONTROL
But nobody knows why
Evidence for gain control in the AL
•These neurons can fire up to100 Hz
•The baseline firing rate is 3-4Hz
Data from Mark Stopfer, Vivek Jayaraman and Gilles Laurent
Honeybee: Galizia’s group
•There seems to be local GABA circuits in the MBs.
•Locust and honeybee circuits are different:
Honeybee 10 times more inhibitory neurons than locust
Let’s concentrate on the locust problem:
How do we design the AL circuit such that it has gain control?
I
i
N
j
N
j IIj
IIij
Ej
IEijI
Ii
I
Ei
N
j
N
j EIj
EIij
Ej
EEijE
Ei
E
fIfwfwtd
fd
fIfwfwtd
fd
E I
E I
1 1
1 1
1
1
Mean field of 4 populations of neurons
inputreceivetheyiSE / inputreceivetheyiS I /
ESi
Ei
E
tfS
tx )(1
)(1
EE NS II NS
ISii
I
tfS
ty )(1
)(1
ESi
Ei
E
tfS
tx )(1
)(2
ESii
I
tfS
ty )(1
)(2
We apply mean field
E
E I
E
E I
Si
N
j
N
j EIj
EIij
Ej
EEij
E
E
Si
N
j
N
j EIj
EIij
Ej
EEijE
E
IfwfwS
F
IfwfwS
1 1
1 1
1
1
Define new set of variables 21 )1( xxgpNX IEIEE 21 )1( yygpNY IEIEI
To obtain the mean field eq.
YYgp
gpXFIY
gp
gpXFgpNY
XYFIYFgpNX
IEIEI
IIIIII
EIEI
IIIIIIEIEI
EEEEIEIEE
)1(
)1(
Where we use 0EEp
constIx ),(* We look for the condition such that
Whose condition is:
*1
*2
*1
*1
*2
*1 ,,
Y
I
IYYIIII
Y
E
EXXEIEI
dudF
FDgp
dudF
FDgp
with21
21)1(
,uu
uu du
d
du
dD and IYXYX EE *
1*2
IYXXYXY II *1
*2
This works if and are linear
EF IF
BUT!
The gain control depends only on the inhibitory connections
The excitatory neurons are not at high spiking frequencies orsilent, but but not very high (3-4) Hz. So0*
2 EYX
ceX
erfXa
XF XE
22 2/21
22)(
SIMULATIONS: 400 Neurons
The gain control condition from the MF can be estimated as 2/EIEIIIII pgpg
A few conclusions:
•Gain control can be implemented in the AL network
•It can be controlled by the inhibitory connectivity. The rest of the parameters are free.
Things to do:
I do not know whether under different odor intensities the AL representation is the same.
Thanks to • Marta Garcia-Sanchez• Loig Vaugier• Thomas Nowotny• Misha Rabinovich
• Vivek Jayaraman• Ofer Mazor• Gilles Laurent