Presented at IEEE Neural Networks Conference, Orlando, FL, Aug.

, 2007

Character Recognition using Spiking Neural

Networks
Ankur Gupta and Lyle N. Long

Abstract A spiking neural network model is used to identify

characters in a character set. The network is a two layered
structure consisting of integrate-and-fire and active dendrite
neurons. There are both excitatory and inhibitory connections
in the network. Spike time dependent plasticity (STDP) is used
for training. It is found that most of the characters are recognized
in a character set consisting of 48 characters.

I. I NTRODUCTION
It is well known that biological neurons use pulses or spikes
to encode information. Neurological research also shows that
the biological neurons store information in the timing of
spikes. Spiking neural networks belong to the third generation
of neural networks and like their biological counterparts use
spikes to represent information flow. They can use spatiotemporal information in communication and computation similar to biological neurons. As they use pulse coding for
information processing, they are much faster than rate coding
which implies some averaging mechanism, and is typically
slower [1] [2]. This doesnt mean though that the rate coding
scheme is never used but it means that pulse coding is used
whenever faster speed is desired [2].
There have been many studies in the past using spiking
neuron models to solve different problems. For example [3]
used spiking neurons for spatial and temporal pattern analysis.
They provided a biologically plausible learning algorithm for
realizing RBFs (Radial Basis Functions), which themselves
are quite powerful in function approximation, pattern classification etc. [4]. In this study, spiking neurons were used to
compute RBFs by storing information in their delays. The time
difference between the pre and the post synaptic spikes was
used to learn these delays.
Panchev et al. [5] [6] [7] proposed a spiking neuron model,
called the ADDS (Active Dendrite and Dynamic Synapse)
model and used it for instructing a robot for navigation and
grasping tasks. They used a mobile Khepera robot [8] equipped
with a camera, gripper, and differential wheels for the real
world tests. The task is to navigate the robot in a restricted
environment with objects of different shape and color such that
it executes specific instructions such as go right or find blue
ball without colliding with other objects in the way.
Baig [9] developed a spatial temporal artificial neuron
model for online cursive handwritten character recognition.
Graduate student, Dept. of Computer Science and Engineering,
Email: [email protected]
Distinguished Professor of Aerospace Engineering and Mathematics,
Email: [email protected]
The Pennsylvania State University, University Park, PA 16802

Their model had the capability to handle spatial-temporal data

sequences continuously. Buonomano et al. [10] proposed a
model for position invariant character recognition by coding
input intensity by relative firing time. They used traditional
backpropagation to learn to discriminate histograms. Jaramillo
et al. [11] used a network of spiking neurons to recognize
short ( 1 sec) spatial-temporal patterns.
The idea of using one spike per neuron to process information was explored by [12]. The order in which the neurons
fired in a particular layer was used as a code. They justified
their assumption by observing the electrophysical data of the
monkey temporal lobe [13] [14] [15]. It was observed that
some neurons of the monkey temporal lobe responded to a face
stimuli with a latency of 80100 ms. After taking into account
that the information has to pass through 10 different stages and
the conduction velocities of the neocortical fibre, it was found
that each stage had less than 10ms for computation. Such rapid
processing presents problem for conventional rate coding if
more than one spike is to be emitted by a cell [12]. Later [16],
also using one spike per neuron, were able to train the network
using spike time dependent plasticity (STDP) to recognize
natural photograph of faces. They used spikeNet [17] for their
simulations.
The present paper attempts to build and train a network
using STDP such that after training it is able to identify
different characters in a character set. The idea is that when the
stimulus (in form of a constant current) is presented as input
to the neurons, the output should be in the form of a series of
spikes i.e. a neuron should respond to a familiar stimulus by
spiking continuously unless another stimulus is presented.
The current paper is organized as follows: The next section
presents the neuron model and the plasticity rule used. The
neural network structure and the learning rule are presented
in the third section. The last two sections present results and
conclusions, respectively.
II. N EURON M ODEL AND P LASTICITY
The neuronal model used is the active dendrite and dynamic
synapse model [6] [18] except that the dynamic synapses have
not been used. We provide a brief introduction to the model in
this section. In this model, a neuron receives input via spike(s)
through a set of synapses and dendrites. The total post-synaptic
current for the synapse i, with weight wi attached to a dendrite
is given by:
di

dIdi (t)
= Idi (t) + Rdi wi (t tif )
dt

(1)

Here tif is the set of pre-synaptic spike times filtered as

Dirac pulses. The time constant di and resistance Rdi defines
the active property of the dendrites as the function of synaptic
weights and are defined as:
di = max |wi |(max min ), |wi | 1

(2)

From the above equation we can see that for high weights,
di is closer to min , whereas for low weights, di is closer to
max . Thus, as the time constant is low for stronger synapses,
we have a earlier and steeper increase of the soma potential
as compared to weaker synapses. The resistance Rdi is given
by:
Rdi =

di
Rm

m
di

m
m i
d

(3)

Here is the neurons firing threshold, Rm is the soma

resistance, and m is the soma time constant. The above
equation for Rdi ensures that the maximum value of the
membrane potential change is proportional to the neurons
firing threshold .
The other influence to an output neuron comes from the
somatic synapses feeding directly or close to the soma.
The equation governing the post-synaptic current for these
synapses is give by:
s

X
dIs (t)
= Is (t) +
wi (t tif )
dt
i

(4)

Finally, combining the contributions from the dendritic

connections and the synapses feeding directly to the soma,
we get the following governing equation for the total soma
membrane potential, um :
dum (t)
= um (t) + Rm (Id (t) + Is (t))
(5)
dt
P
where, Id (t) = i Idi (t) is the total dendritic current, m
is the soma time constant, and Rm is the soma resistance.
When the membrane potential reaches the threshold value ,
it produces a spike, and immediately after that the membrane
potential is reset to a value ureset = 1mV . After this event,
the membrane potential recovers to the resting potential value.
The learning for both, the synapses attached to active
dendrites and the somatic synapses takes place using spike
time dependent plasticity (STDP) [19] [20] [21] [22] [23]
rules. STDP emphasizes that synaptic plasticity depends on the
precise timing and temporal order between the pre and post
synaptic spikes. It serves to address some of the major shortcomings of the Hebbian based learning such as competition
between synapses and accidental pre/post synaptic correlation
activity [22].
The general STDP learning rule for the synapses is given
by:
(
t
if t < 0
A+ e +
(6)
w =
t
A e if t > 0
m

Where, t = tpre tpost . Here the goal is that weights are

changed such that the next (after this training iteration) postsynaptic spike occurs closer to the pre-synaptic spike. The
weights are then changed according to the relaxation rule:

wold + w(wmax wold ) if w 0
wnew =
(7)
wold + w(wold wmin ) if w < 0
Here, is the learning rate. For excitatory synapses wmin = 0
and wmax = 1, whereas for inhibitory synapses wmin = 1
and wmax = 0.
III. T EST P ROBLEM AND N ETWORK S TRUCTURE
In order to test the network, a character set, shown in
figure 1 consisting of 48 characters is used. Each character
is represented by an array of 3X5 pixels. Integrate and fire
neurons [24] with constant input current are used as input
neurons. If the pixel is on, a constant current is supplied
to the corresponding neuron, whereas if the pixel is off, no
current is supplied to that particular neuron. Figure 2 shows
a typical representation of the character A by the input
neurons.
The various parameters used in the network are defined
in the appendix. The number of input neurons is equal to
the number of pixels in the image. Thus there are 15 input
neurons in the present case. The number of output neurons is
the number of characters to be trained. There are two layers in
the network. The first layer consists of simple leaky integrate
and fire neurons which receive constant or zero input current,
corresponding to on or off states of the input pixels. The
next layer is the layer of active dendrite neurons, each of which
is connected to all of the neurons in the previous layer. Finally,
each of the output layer neuron is connected to every other
output neuron via inhibitory lateral connections. These lateral
connections reflect the competition among the output neurons.
IV. R ESULTS
For initial testing, the network was trained using only four
characters (A, B, C, and D). There were 15 input
neurons and 4 output neurons for this case. The training
parameters used are described in the appendix. The weights
were initialized to random values between 0.5 and 1.0, so
that all the output neurons spike at the first training iteration.
Each character was presented one at a time sequentially during
the training process. When the Forbenius norm of the weight
change matrix was very small, it was assumed that the network
is fully trained and no further significant learning is possible.
For this simple case, the Forbenius norm of the weight change
matrix was 103 after 100 epochs (an epoch means one full
presentation of all the four characters), and thus the training
was stopped at this point. The simulation time step was 0.2ms.
Figures 3(a), 3(b), 3(c), 3(d) show the output of all the
four neurons when the characters A, B, C, and D are
presented as inputs respectively after training. We can see that
only one neuron responds to a particular character, by spiking
continuously with spike-time interval of 100ms, whereas the
other neurons are below the threshold value of 10mV . Note

Fig. 1.

Fig. 2.

Character set used

Representation of A

(a) Character A

(b) Character B

(c) Character C

(d) Character D

Fig. 3.

Output when each character is presented individually

that here the current of the input neurons is set to a constant

value such that the neurons spike after every 100ms interval.
It was observed that the minimum spike-time interval allowed
is roughly 80ms, for the network to respond correctly. This
is because this is roughly the time required for the neurons to
reach the resting membrane potential.
Figure 4 shows the output of the network when the
characters are presented one by one in the order C, D, A,
B, C, D, ... and so on. A new character was presented
1ms before every 300ms. We can again see that only a
particular neuron responds to a particular character by spiking
continuously unless the next character is presented, when
another neuron starts spiking, and so on. Figures 5(a) and
5(b) show the weights of each of the connections before and
after training respectively.
The next test case was trying to train the full 48 character
set in figure 1. The network here consisted of 15 neurons
in the input layer, and 48 in the output layer. The network
structure was the same as previous except that there were

Fig. 4. Output when characters are presented in following order: C, D,

A, B, C, D, ...

more connections as the number of output neurons was 48.

The network was trained for 200 epochs. During this training
process each of the 48 characters was presented sequentially
until the Forbenius norm of the weight change matrix was
very small. Figure 6 shows the variation of the Forbenius
norm with the number of training epochs. It decreases roughly

(a) Before training

(b) After training

Fig. 5.

Weight distribution

linearly (log scale) with number of epochs.

Most of the characters (43) were recognized uniquely in
the sense that either a unique neuron fired or the firing
times of the same neuron were different. A single neuron
responded with same firing times for the characters M, U,
and W. Strikingly, all have the same number of pixels (11) in
there character representations and look very similar. Similarly,
characters J and 3, each having 10 pixels, also had nonunique representations. Rest of the characters had a unique
representation.
Figure 7 shows the soma potentials of all the 48 neurons
after training on presentation of a single character. Each
symbol along with the color represents a unique neuron. We
can see that only one neuron ( symbol with cyan color),
which is the winner, spikes, and the rest remain below the
threshold value of 10mV . Similar observations were made
with rest of the characters.

Fig. 7. Soma potentials for the case of 48 characters: Only one is the winner

change was observed. Most of the characters were recognized

when the network was trained using a character set of 48
characters. Preliminary results look encouraging and more
complex problems such as image recognition will be tried in
the future.
ACKNOWLEDGMENT
This research was supported by ONR Grant No. N0001405-1-0844.
A PPENDIX I
RUNNING PARAMETERS
Rm = 80, = 10mV , s = 2ms, m = 30ms, min =
2ms, max = 30ms, = 0.1, decay = 0.05, A+ = 0.1,
A = 0.105, + = 1ms, = 1ms, time step = 0.2ms
R EFERENCES

Fig. 6. Variation of Forbenius norm of the weight change matrix with no.
of epochs

V. C ONCLUSION
A two layered spiking neural network was used to identify
characters in a character set. STDP was used to train the
network. The network was trained until no significant weight

[1] W. Gerstner, R. Kempter, J. L. van Hemmen, and H. Wagner, Hebbian

Learning of Pulse Timing in the Barn Owl Auditory System, ser. Pulsed
Neural Networks, W. Maass and C. M. Bishop, Eds. Berlin, Germany:
MIT-press, 1999.
[2] S. Thorpe, A. Delorme, and R. van Rullen, Spike based strategies for
rapid processing, Neural Networks, vol. 14, pp. 715726, 2001.
[3] T. Natschlager and B. Ruf, Spatial and temporal pattern analysis via
spiking neurons, Network: Comput. Neural Syst., vol. 9, pp. 319332,
1998.
[4] S. Haykin, Neural Networks: A Comprehensive Foundation. Upper
Saddle River, NJ: Prentice-Hall, 1999.

[5] C. Panchev and S. Wermter, Spike timing dependent synaptic plasticity:

from single spikes to spike trains, Neurocomputing, vol. 58-60, pp. 365
371, 2004.
[6] C. Panchev, Spatio-temporal and multimodal processing in a spiking
neural mind of a robot, Ph.D. dissertation, Univeristy of Sunderland,
United Kingdom, May 2005.
[7] C. Panchev and S. Wermter, Temporal sequence detection with spiking
neurons: Towards recognizing robot language instruction, Connection
Science, vol. 18, pp. 122, 2006.
[8] [Online]. Available: http://www.k-team.com/
[9] A. R. Baig, Spatial-temporal artificial neurons applied to online cursive
handwritten character recognition, in European Symposium on Artificial
Neural Networks, April 2004, paper Bruges (Belgium), pp. 561566.
[10] D. V. Buonomano and M. M. Merzenich, A neural network model of
temporal code generation and position invariant pattern recognition,
Neural Computation, vol. 11, pp. 103116, 1999.
[11] S. Jaramillo and F. Pereira, Recognition of dynamic patterns
with a network of spiking neurons. [Online]. Available:
http://www.cs.unm.edu/ sjara/docs/Jaramillo and Pereira CSSS2002.pdf
[12] R. van Rullen, J. Gautrais, A. Delorme, and S. Thorpe, Face processing
using one spike per neurone, Biosystems, vol. 48, pp. 229239, 1998.
[13] C. J. Bruce, R. Desimone, and C. G. Gross, Visual properties of neurons
in a polysensory area in superior temporal sulcus of the macaque,
Journal of Neurophysiology, vol. 46, pp. 369384, 1981.
[14] D. I. Perrett, E. T. Rolls, and W. Caan, Visual neurons responsive to
faces in the monkey temporal cortex, Experimental Brain Research,
vol. 47, pp. 329342, 1982.

[15] D. A. Jeffreys, Evoked potential studies of face and object processing,

Visual Cognition, vol. 3, pp. 138, 1996.
[16] A. Delorme and S. J. Thorpe, Face identification using one spike per
neuron: resistance to image degradations, Neural Networks, vol. 14, pp.
795804, 2001.
[17] [Online]. Available: http://www.sccn.ucsd.edu/ arno/spikenet/
[18] C. Panchev, S. Wermter, and H. Chen, Spike-timing dependent competitive learning of integrate-and-fire neurons with active dendrites,
in Proceedings of the International Conference on Artificial Neural
Networks, August 2002, pp. 896901.
[19] H. Markram, J. Lubke, M. Frotscher, and B. Sakmann, Regulation of
synaptic efficacy by coincidence of postsynaptic aps and epsps, Science,
vol. 275, pp. 213215, 1997.
[20] G. Q. Bi and M. M. Poo, Synaptic modifications in cultured hippocampal neurons: dependence on spike timing, synaptic strength, and
postsynaptic cell type, J. Neurosci., vol. 18, pp. 10 46410 472, 1998.
[21] Y. Dan and M. Poo, Spike time dependent plasticity of neural circuits,
Neuron, vol. 44, pp. 2330, 2004.
[22] S. Song, K. D. Miller, and L. F. Abbott, Competitive hebbian learning through spike-timing-dependent synaptic plasticity, Nature Neuroscience, vol. 3, no. 9, pp. 919926, 2000.
[23] B. Berninger and G. Q. Bi, Synaptic modification in neural circuits: a
timely action, BioEssays, vol. 24, pp. 212222, 2002.
[24] W. Gerstner and W. M. Kistler, Spiking Neuron Models: Single Neurons,
Populations, Plasticity. Cambridge University Press, 2002.