ON-CENTER/OFF-SURROUND NEURAL NETWORK MODEL

FOR OLFACTORY ATTENTION

Zu Soh

Department of Biotechnology, Osaka University, 2-1 Yamadaoka, Shuita, Osaka, Japan

Toshio Tsuji

Department of System Cybernetics, Hiroshima University, 1-4-1 Kagamiyama, Higashi-Hiroshima, Japan

Noboru Takiguchi

Division of Material Sciences, Kanazawa University, Kakuma, Kanazawa, Ishikawa, Japan

Hisao Ohtake

Department of Biotechnology, Osaka University, 2-1 Yamadaoka, Shuita, Osaka, Japan

Keywords:

Olfactory system, Attention, Neural network model.

Abstract:

Our research group has found behavioral evidence that an attention function exists in the olfactory system

similarly to in the visual and auditory systems. In this paper we propose a neural network model that accounts

for olfactory attention based on macroscopic neural connections. Speciﬁcally, on-center/off-surround connec-

tions were assumed to be involved in the attention process in accordance with our hypothesis of an attention

window that extracts local activity. The model employs glomerular activity patterns as its input, and compares

them with stored patterns focusing on their local activity. The model also can shift and change the attention

window with respect to learning. From the simulation results, we conﬁrmed that the model can account for

the results of a behavioral experiment on olfactory attention in mice.

1 INTRODUCTION

Attention is an important cognitive function for ﬁl-

tering out irrelevant information and extracting use-

ful information from a noisy environment. Atten-

tion thus enables efﬁcient information processing in

the case of limited computational capacity (Dayan

et al., 2000). The mechanisms of visual and auditory

attention have been studied over previous decades

(Broadbent, 1958) from the behavioral level to the

neural level. These studies suggest that selective at-

tention can be caused by modulation of the ampli-

tude of neural activity evoked by stimuli (Hillyard

et al., 1998). As physiological mechanisms have been

gradually elucidated, their mathematical models have

begun to be applied to robotics (Vijayakumar et al.,

2001; Ruesch et al., 2008). In contrast to these devel-

opments for visual and auditory systems, to the best

of our knowledge, the existence of an attention mech-

anism in olfactory systems has only recently begun to

be investigated.

As natural odors are generally composed of a

complex mixture of volatile compounds (odorants),

of which more than 400,000 types (Mori et al., 2006)

exist, focusing on part of them should be an efﬁcient

means of recognizing odor. To determine whether an

attention function also exists in the olfactory system,

our research group performed an odor discrimination

experiment on mice (Takiguchi et al., 2008). The

experiment provided evidence that mice can mem-

orize and discriminate odors by focusing on a sub-

set of odorants comprising an odor. Furthermore,

when a difﬁcult discrimination task was presented,

the mice slowly modiﬁed the attention subset through

their learning experience. As visual and auditory sys-

tems can quickly switch their attention to different ob-

183

Soh Z., Tsuji T., Takiguchi N. and Ohtake H..

ON-CENTER/OFF-SURROUND NEURAL NETWORK MODEL FOR OLFACTORY ATTENTION.

DOI: 10.5220/0003680301830189

In Proceedings of the International Conference on Neural Computation Theory and Applications (NCTA-2011), pages 183-189

ISBN: 978-989-8425-84-3

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Odor

boxes

Mouse

E1 E2

Odor outlet

Water-feeder

Figure 1: Y-maze.

jects, from this viewpoint the attention function in the

olfactory system is different from that in the visual

and auditory systems.

Although an attention function was found at the

behavioral level, its mechanism at the neural level is

not clear. As current technology does not allow the

exhaustive measurement and interpretating of neural

activity, building a mathematical model is a perspec-

tive approach to this problem. Most previous olfac-

tory models were built with the aim of accounting

for neural dynamics (Cleland and Sethupathy, 2006)

and background segmentation (Li and Hertz, 2000);

however, a model to account for the attention mech-

anism has not yet been proposed. Against this back-

ground, we previously proposed an attention model

that could predict perceptual similarity between odors

(Soh et al., 2009); however, several assumptions em-

ployed in the model were not realistic, and the model

could not explain shifts and changes in attention. In

this paper, we propose an olfactory attention model

taking the macroscopic neural connections in the ol-

factory system into account. The model employs neu-

ral activity evoked on the glomeruli as its input and

predicts the odor discrimination ability of mice re-

sulting from their attention function. In addition, we

demonstrate that the discrimination ability predicted

by the model has a similar tendency to that observed

in experiments.

2 EXPERIMENT ON

OLFACTORY ATTENTION

To elucidate the attention mechanism in the olfac-

tory system, our research group performed an odor

discrimination experiment on mice (Takiguchi et al.,

2008) using different mixtures of three types of odor-

ant: isoamyl acetate (IA), ethyl butyrate (EB), and

Citral (Ci). In the experiment, the subject mice were

prompted to discriminate between a water-rewarded

odor [IA, Ci, EB] and odors composed of different

Correct

discrimination rate

Later 12 trialsFirst 12 trials

Rewarded odor [IA,EB,Ci]

100

[IA,EB] [IA,Ci] [EB,Ci] [IA] [EB] [Ci]

Figure 2: Results of odor discrimination experiment.

combinations. In this section we review the procedure

and experimental results.

The experiment employed a Y-maze assay, which

uses a Y-shaped channel as shown in Figure 1. At the

end S, there is a small chamber with a gate (starting

box) separated from the channel to trap the subject.

The other ends E1 and E2 have attached odor boxes,

which transmit odors into the channel, and water feed-

ers to provide the reward. During the experiment the

subject mice were prohibited from drinking water ex-

cept as a reward.

A trial started with setting the rewarded odor [IA,

Ci, EB] and an unrewarded odor in the two odor

boxes. The subject was then placed in the starting

box. When the gate opened, the subject could run

through the channel and choose one of the branches

of the channel depending on the response to the odor.

The subject was allowed to drink water from the water

feeder next to the rewarded odor, otherwise it was re-

turned to the starting box immediately. This trial was

repeated 24 times in a day, referred to as one session.

After the subjects were trained to form an asso-

ciation between the rewarded odor [IA, Ci, EB] and

water in the ﬁrst several sessions, the subjects were

required to discriminate between the rewarded odor

and unrewarded odors. Figure 2 shows the average

discrimination rate of eight subjects in the ﬁrst 12 tri-

als and the following 12 trials in a session. The dis-

crimination rate of 46.9% against odor [IA, EB] in

the ﬁrst 12 trials implies that the subjects considered

odor [IA, EB] as the rewarded odor, since a discrimi-

nation rate of 50% would be expected as a result of

random selection. In the following 12 trials, how-

ever, the subjects became more able to discriminate

between these odors since the average discrimination

rate increased to 66.7%. This suggests that when the

association between water and the rewarded odor [IA,

Ci, EB] was made, the subjects only paid attention to

the odorants IA and EB. Subsequently, the attention

of the mice changed, enabling them to perform cor-

NCTA 2011 - International Conference on Neural Computation Theory and Applications

184

Olfactory

bulb

Olfactory

epithelium

Piriform

cortex

Odorants

Olfactory epithelium Olfactory bulb Piriform cortex

Receptor

neuron

Glomeruli Pyramidal cell

䊶䊶䊶

Mitral cell Granular cell

Excitatory synapse

Inhibitory synapse



Geranyl acetate

Glomerular activity pattern

Figure 3: Structure of the olfactory system.

rect discrimination. These results thus conﬁrmed the

attention ability of the olfactory system.

Although this experiment illustrated the occur-

rence of incorrect discrimination caused by attention

ability, considering the large number of odorants that

exist, paying attention to a few important odorants

comprising odors is much more efﬁcient for identi-

fying or discriminating between odors. Despite its

importance, the mechanism of attention at the neu-

ral level has not been investigated. To explain the at-

tention ability of the olfactory system, a window that

extracts local neural activity evoked by odors was as-

sumed (Takiguchi et al., 2008). Discrimination be-

tween odors is then performed using only the neural

activity included in this window. However, this hy-

pothesis has not been validated.

3 STRUCTURE OF THE

OLFACTORY SYSTEM

As the mechanism of attention has not been eluci-

dated in the biological ﬁeld, partly because of the dif-

ﬁculty of exhaustively measuring and interpretation

of neural activity, in this paper, we propose a possible

attention model from an engineering approach based

on the neural structure of the olfactory system (Mori

et al., 2006; Heimer, 1968) and the above hypothe-

sis (Takiguchi et al., 2008). In this section we brieﬂy

review the structure of the olfactory system.

Figure 3 shows the basic structure of the olfac-

tory system of mice, which consists of three parts: re-

ceptor neurons, the olfactory bulb, and the piriform

cortex. Receptor neurons are distributed on the sur-

face of the nasal chamber, expressing a single recep-

tor protein from among thousands of different vari-

eties (Buck and Axel, 1991); each neuron is activated

by a speciﬁc group of odorants and sends signals to

the olfactory bulb.

The olfactory bulb mainly consists of glomeruli,

mitral cells, and granular cells. A glomerulus is a

round cluster of axon terminals accumulated from

receptor neurons. The activity patterns evoked on

glomeruli are odor-speciﬁc ((Mori et al., 2006; John-

son and Leon, 2000), shown in the lowest part of Fig-

ure 3). A mitral cell is an excitatory neuron that re-

ceives the output from a glomerulus. Granular cells

are inhibitory neurons sending inhibitory signals to

the mitral cells. Although mitral cells and granular

cells appear to form complex connections, it has been

suggested that they form an on-center/off-surround

circuit in which neighboring mitral cells excite each

other but distant cells inhibit each other (Grossberg,

1976). The mitral cells also transfer signals to the

pyramidal cells in the piriform cortex, which then

send signals back to the granular cells in the olfactory

bulb and indirectly inhibit the mitral cells.

On-center/off-surround connections in the olfac-

tory bulb are well-known neural connections found

in sensing systems and typically perform contrast en-

hancement. In this paper, we consider that these con-

nections have an important role in generating a win-

dow of attention to extract neural activity.

4 PROPOSED MODEL

Since several experimental results and mathematical

models have suggested that the olfactory bulb has

the functions of input normalization and contrast en-

hancement (Grossberg, 1976; Cleland and Sethupa-

thy, 2006), we modeled the attention function as an in-

teraction between these two functions. However, the

interconnections involved in the olfactory system can

evoke complex dynamics that preventus from analyz-

ing attention mechanisms; we thus designed a simple

model that makes it possible to focus on spatial neu-

ral activity patterns taking the macroscopic connec-

tions between neurons into account. In this section,

the structure of each layer and the parameter settings

of the proposed model are described.

4.1 Structure of the Proposed Model

Figure 4 shows the structure of the proposed model,

which consists of three layers: the input layer, the ol-

factory bulb layer, and the output layer.

The input layer carries the activity patterns evoked

on the glomeruli by odorants o

mixed in an odor O =

ON-CENTER/OFF-SURROUND NEURAL NETWORK MODEL FOR OLFACTORY ATTENTION

185

...

Neuron

units

Glomerular

activity patterns

Olfactory bulb

layer

Output layer

wjp

㪅㪅㪅

Input layer

Inhibitory input

Output

Figure 4: Structure of the proposed model.

,...,o

]. In this paper, we used the activity pat-

terns measured from actual organisms obtained from

an online database (http://gara.bio.uci.edu/). The pro-

vided activity patterns were originally in an image ﬁle

format. To convert each image ﬁle into an input, the

image is divided into L = 1805 lattice squares approx-

imately corresponding to the number of glomeruli on

a mouse’s olfactory bulb. Each of the lattice squares

is then converted into a value ranging from 0 to 1 de-

pending on the color in the lattice, which corresponds

to the activity strength. The activity evoked on the jth

lattice square by odorant o

is thus denoted as a

, and

the activity pattern is the vector a

∈ R

1×L

The olfactory bulb layer consists of L neuron

units. They receive inputs from the input layer as

well as inhibitory inputs from the output cortex layer.

To realize the functions of normalization and con-

trast enhancement, this layer employs a neural net-

work model whose structure was previously proposed

(Grossberg, 1976), given as

∑

i=1

, (1)

˙u

= −Tu

+ (B− u

− u

(

∑

′

6= j

′

− H

), (2)

where I

′

is the input from the input layer, which is

the summation of activity strengths evoked by odor-

ants, u

is the activity of the jth neuron unit, T, B, and

L are constant coefﬁcients, H

is the inhibitory input

adjusted by the output layer, and w

′

is the weight co-

efﬁcient representing the on-center/off-surround con-

nections as follows;

′







W, d < D

−W, D

< d < D

0, d > D

. (3)

Here, W is a constant, d is the distance between the

units, calculated from their corresponding locations

on the glomeruli, and the coefﬁcients D

and D

de-

note the maximum distances of the on-center excita-

tory connections and off-surround inhibitory connec-

tions that exist, respectively. u

at the equilibrium

point (˙u

= 0) is then

T +

∑

<d<D

−

∑

d<D

− H

; (4)

thus, u

is normalized by the total input of neuron

units in the region D

< d < D

and is enhanced by

those in the region d < D

Finally, the output layer calculates the sum of the

outputs from the olfactory bulb layer as follows:

∑

f(u

), (5)

f(x) =

1+ exp[−g(x− θ

)]

, (6)

where w

is the weight coefﬁcient, and g and θ

are

the gain and threshold constants of the sigmoid func-

tion, respectively. Because the output u

is the cal-

culated inner product between the stored pattern and

input pattern when an activity pattern is stored in w

is deﬁned as the correspondence with the stored

odor.

4.2 Proposed Attention Process

To implement the hypothesis that a window extracts a

local activity pattern for attention, the weight coefﬁ-

cient w

is determined as follows:



0, u

) < θ

1/P, u

) ≥ θ

, (7)

where O

represents a rewarded odor such as [IA,

Ci, EB] used in the odor discrimination experiment.

) is then the activity pattern of the olfactory bulb

layer evoked by odor O

, θ

is the threshold constant

for u

, and P is the total number of neuron units in

the olfactory bulb layer whose output is greater than

. This allows the output layer to compare the activ-

ity patterns by focusing only on the part activated by

odor O

, which is expressed in the window of atten-

tion.

As observed in the odor discrimination experi-

ment, attention can cause incorrect discrimination. In

this case, the window has to be changed. To model

this function, we apply the inhibitory input H

to the

olfactory bulb layer when an odor different from the

rewarded odor generates a high output for u

. The

inhibitory input H

is thus determined as follows:



βu

, u

< θ

and u

) ≥ θ

0, otherwise

, (8)

NCTA 2011 - International Conference on Neural Computation Theory and Applications

186

[IA] [EB] [Ci] [Li]

Figure 5: Glomerular activity patterns evoked by the odor-

ants in odor discrimination experiments.

where θ

is the threshold constant for u

, and β is a

gain coefﬁcient used to determine the strength of H

This conﬁguration generates new activity by turning

off inhibition from off-surround connections, and the

activity produces a new window in accordance with

Equation 7.

5 SIMULATION

5.1 Simulation Procedure

The model was validated by performing two sets of

discrimination experiments. First, the parameters of

the model were adjusted manually in accordance with

the experimental results when a combination of IA,

EB, and Ci was used. The parameters were then vali-

dated by comparison with another set of experimental

results using the odorants IA, linaool (Li), and Ci. Fi-

nally, the ability to change the attention window was

also investigated.

As the input to the model, we used mea-

sured glomerular activity patterns evoked by the

odorants in the odor discrimination experiment.

These were obtained from the database website

http://gara.bio.uci.edu/,where the activity patterns are

provided as unrolled maps of the spherical surface of

the olfactory bulb as shown in Figure 5. However,

since the activity pattern for odorant Ci was not in the

database, we used the activity pattern predicted from

a glomerular activity prediction model (Soh et al.,

2011). Figure 5 shows the activity patterns for each

odorant.

A comparison between the simulation and exper-

imental results was carried out by deﬁning a metric

corresponding to the correct discrimination rate ob-

tained from the experiment on mice described in Sec-

tion 2. As mentioned in the previous section, the out-

put u

represents the correspondence with the stored

pattern in weight coefﬁcient w

p; we thus used the

following metric to represent the correct discrimina-

tion rate:

[IA,EB,Ci] [IA,EB] [IA,Ci]

[EB,Ci]

[IA]

[Ci]

[EB]

Figure 6: Activity patterns on the olfactory bulb layer for

each odorant obtained by simulation.

Attention

window

(wjp)

Filtered

[IA,EB,Ci]

Filtered

[IA,EB]

Filtered

[Ci]

Figure 7: Attention window and activity patterns ﬁltered by

the window.

r =

)

) + u

)

, (9)

where u

) is the output of the rewarded odor and

) is that of the unrewarded odor. The param-

eters were thus adjusted to ﬁt r with the correct dis-

crimination rate.

5.2 Results

First, the parameters were adjusted. Similarly to in

the odor discrimination experiment, odor [IA, EB, Ci]

was used as the rewarded odor. As a result of simu-

lation, the activity patterns of the olfactory bulb layer

shown in Figure 6 were obtained. In addition, Figure

7 shows the attention window, represented by white

spots generated by the model, and examples of ac-

tivity patterns on the olfactory bulb layer ﬁltered by

the attention window. This ﬁgure demonstrates how a

comparison between odors can be made by focusing

on the window generated by the model.

Figure 8 shows the values of r obtained from the

model with adjusted parameters and the correct dis-

crimination rate against each odorant. This ﬁgure

conﬁrmed that r and the correct discrimination rate

have a similar tendency, for example, the lowest cor-

rect discrimination rate was observed when odor [IA

EB] was the unrewarded odor. These results con-

ﬁrmed that the model focused attention on the activity

ON-CENTER/OFF-SURROUND NEURAL NETWORK MODEL FOR OLFACTORY ATTENTION

187

Correct discrimination

rate %

0.4

0.6

0.8

[IA,EB] [IA,Ci] [EB,Ci] [IA] [EB] [Ci]

100

Simulation result of the model

Average correct discrimination rate of mice

Figure 8: Comparison of attention between model and mice.

(rewarded odor: IAEBCi).

Correct discrimination

rate %

0.2

0.4

0.6

0.8

[IA,Li] [IA,Ci] [Li,Ci] [IA] [Li] [Ci]

100

Simulation result of the model

Average correct discrimination rate of mice

Figure 9: Comparison of attention between model and mice.

(rewarded odor: IALiCi).

evoked by [IA EB], as observed in the odor discrim-

ination experiment. The adjusted parameters were as

follows.

• Olfactory bulb layer

T = 0.1, B = 2.0,

= 3.0, D

= 10.0, W = 0.012

• Output layer

g = 0.03,θ

= 0.7

Using the adjusted parameters, a simulation was

performed with another odor set, where odor [IA, Li,

Ci] was used as the rewarded odor. Figure 9 shows

the values of r obtained from the model and the cor-

rect discrimination rate of mice against each odorant.

From the ﬁgure, we can conﬁrm that the proposed

model and the mice have a similar tendency that both

focused attention on odorant [IA].

We then tested the ability to change the attention

window by applying an inhibitory input H

to the ol-

factory bulb layer. Since the discrimination ability of

the mice improvedin the later 12 trials as described in

Section 2, we also investigated whether the value of r

increases when the attention window is changed. For

this simulation, we set [IA, EB, Ci] as the rewarded

odor. Figure 10 shows the changes in the attention

window with increasing strength of the inhibitory in-

put (parameter β). From the ﬁgure, we can conﬁrm

that the window changed with increasing β. We also

conﬁrmed that when β was increased from 0 to 0.3,

r increased from approximately 0.48 to 0.72. These

results conﬁrmed the ability to change the attention

=0.2

=0.3=0.1

β β

Figure 10: Changes in attention window (white spots).

window, as observed in the odor discrimination ex-

periment on mice.

6 CONCLUSIONS

We proposed an attention model for the olfactory sys-

tem assuming the existence of an attention window

generated by normalization and contrast enhancement

processes in the olfactory bulb. A possible learning

mechanism was also proposed in which an inhibitory

signal is applied to the olfactory bulb, which modiﬁes

the attention window. Although the model only con-

sidered the macroscopic structure of the olfactory sys-

tem, a comparison with the correct discrimination rate

of mice conﬁrmed the attention ability of the model.

The correct discrimination rate can also be inter-

preted in terms of the perceptual characteristics of

mice as the Y-maze experiment is a type of sensory

evaluation assay. The proposed model can thus be ap-

plied as an artiﬁcial sensory evaluation method. As a

future work we are planning to increase the odorant

set to validate and improve the model. We are also

considering applying the model to predict the percep-

tual characteristics of human olfaction.

ACKNOWLEDGMENTS

This work was supported by Grant-in-Aid for

Japan Society for the Promotion of Science Fellows

(23·10869).

REFERENCES

Broadbent, D. E. (1958). Perception and Communication.

New York: Pergamon.

Buck, L. and Axel, R. (1991). A novel multigene family

may encode odorant receptors: a molecular basis for

odor recognition. Cell, 65:175–187.

Cleland, T. A. and Sethupathy, P. (2006). Non-

topographical contrast enhancement in the olfactory

bulb. BMC Neurosci, 7:7.

NCTA 2011 - International Conference on Neural Computation Theory and Applications

188

Dayan, P., Kakade, S., and Montague, P. R. (2000). Learn-

ing and selective attention. Nature Neuroscience,

3:1218–1223.

Grossberg, S. (1976). Adaptive pattern classiﬁcation and

universal recoding: I. parallel development and cod-

ing of neural feature detectors. Biol. Cybern., 23:121–

134.

Heimer, L. (1968). Synaptic distributions of centripetal

and centrifugal nerve ﬁbres in the olfactory system of

the rat. an experimental anatomical study. Journal of

Anatomy, 103:413–432.

Hillyard, S. A., Vogel, E. K., and Luck, S. J. (1998).

Sensory gain control (ampliﬁcation) as a mechanism

of selective attention: Electrophysiological and neu-

roimaging evidence. Philos Trans R Soc Lond B Biol

Sci., 353:1257–1270.

Johnson, B. A. and Leon, M. (2000). Modular represen-

tations of odorants in the glomerular layer of the rat

olfactory bulb and the effects of stimulus concentra-

tion. J Comp Neurol, 422:496–509.

Li, Z. and Hertz, J. (2000). Odour recognition and segmen-

tation by a model olfactory bulb and cortex. Network,

11:83–102.

Mori, K., Takahashi, Y. K., Igarashi, K. M. ., and Yam-

aguchi, M. (2006). Maps of odorant molecular fea-

tures in the mammalian olfactory bulb. Phys. Rev.,

86(2):409–433.

Ruesch, J., Lopes, M., Bernardino, A., Hornstein, J.,

Santos-Victor, J., and Pfeifer, R. (2008). Multimodal

saliency-based bottom-up attention a framework for

the humanoid robot icub. IEEE International Confer-

ence on Robotics and Automation, pages 962–967.

Soh, Z., Tsuji, T., Takiguchi, N., and Ohtake, H. (2009).

Neuro-based olfactory model for artiﬁcial organolep-

tic tests. Artiﬁcial Life and Robotics, 14:474–479.

Soh, Z., Tsuji, T., Takiguchi, N., and Ohtake, H. (2011). An

artifcial neural network approach for glomerular ac-

tivity pattern prediction using the graph kernel method

and the gaussian mixture functions. Chem. Sens., page

in press.

Takiguchi, N., Okuhara, K., Kuroda, A., Kato, J., and

Ohtake, H. (2008). Performance of mice in discrimi-

nation of liquor odors: Behavioral evidence for olfac-

tory attention. Chem. Sens., 33:283–290.

Vijayakumar, S., Conradt, J., Shibata, T., and Schaal,

S. (2001). Overt visual attention for a humanoid

robot. IEEE/RSJ International Conference on Intel-

ligent Robots and Systems, 4:2332–2337.

ON-CENTER/OFF-SURROUND NEURAL NETWORK MODEL FOR OLFACTORY ATTENTION

189