Pulsed Neural Networks and their Application

Daniel R. Kunkle Chadd Merrigan

[email protected] [email protected]
http://www.rit.edu/~drk4633/ http://www.rit.edu/~ccm2451/

Computer Science Dept.

College of Computing and Information Sciences
Rochester Institute of Technology
Rochester, NY, 14623

February, 2002

Abstract networks. Neurons have been found in the primate brain

that respond selectively to complex visual stimuli after as
Pulsed neural networks are networks of spiking few as 100-150 ms after the stimulus was presented. This
neurons, which represent an entirely new class of information must have passed through approximately 10
artificial neurons. Here we present an overview layers of processing between the initial photoreceptors
of pulsed neural networks, including the
and the neurons that selectively respond to the stimuli.
structure, function and available training
Given this, it was argued that each individual processing
mechanisms for networks of spiking neurons.
We highlight differences between this model, stage would have only 10 ms to complete and that this
“first generation” threshold gates, and “second amount of time is insufficient for rate coding as a means
generation” sigmoid activation gates, and we of information passing (Thorpe et al., 2001).
examine current research into pulsed neural Another information coding scheme must therefore be at
networks and the use of temporal information in work in the brain, one that may be useful in artificial
neural processing. Lastly, we summarize our neural networks. Codes based the temporal relationship
own research toward the end of using pulsed between the firing of neurons is a promising alternative.
neural networks to identify computer users by Using such codes it is possible to transmit a large amount
the cadence of their keystrokes.
of data with only a few spikes, as few as one or zero for
each neuron involved in the specific processing task
1 INTRODUCTION (Thorpe et al., 2001). It is the use and effects of temporal
information coding in artificial networks that will be
Artificial neural networks, being based on the workings of examined here.
biological neural networks, can be expected to draw
inspiration from advances in neurophysiology and related Section 2 comprises an overview of artificial spiking
fields. Traditionally, it was believed that neuron neurons, including the main categories of artificial
communicated information in their mean firing rate. In neurons, a brief overview of the biological basis of the
other words, one neuron would receive information from spiking neuron model and the main benefits and qualities
another by “counting” the number of spikes from that of using artificial spiking neurons.
neuron over some extended period of time and In section 3 the model of an artificial spiking neuron is
determining the mean time between firings. Specifically, examined in depth, detailing its structure and functioning,
a shorter time period implies a higher activation. This and introducing the common mathematical model.
type of information coding is known as rate coding and
has been one of the dominant tools for measuring neuron Section 4 presents methods for training spiking neurons
activity over the past 80 years of neurological study and pulsed neural networks, allowing them to learn and
(Gerstner, 1999). Correspondingly, a majority of artificial act on temporally encoded data.
neural network models have used rate coding, in the form Finally, sections 5 and 6 explore possible applications of
of real number values representing an activation level. pulsed neural nets, including some original work by the
Recently, new discoveries and advances in authors aimed at identifying computer users by the
neurophysiology have encouraged new explorations in cadence of their keystrokes.
alternative information coding schemes in artificial neural
2 OVERVIEW OF SPIKING NEURONS

2.1 Artificial Neuron Generations

Wolfgang Maass (Maass, 1997) delineates past and
current artificial neural network research into three
generations and makes the following observations.
The first generation is based on the McCulloch-Pitts
neuron (also known as a perceptron or a threshold-gate)
as the basic computation unit. Models of the first
generation, such as the multi-layer perceptron, use digital
input and output, usually binary or bipolar. Any boolean
function can be computed by some multi-layer perceptron
with a single hidden layer.
The second generation is based on computation units
(neurons) that use an activation function of a continuous
set of possible output values. Commonly, these activation
functions are the sigmoid, f(x) = 1 / 1 + e- σx, or the
hyperbolic tangent, f(x) = (1 – e-2x) / (1 + e-2x). Second
generation neural networks, like first generation networks,
can compute arbitrary boolean functions (after using a
threshold). Second generation networks can compute Figure 1: Simultaneous recordings (over 4 seconds) of
certain boolean functions with fewer neurons than first the firing times of 30 neurons from monkey striate
generation neurons. Also, second generation networks cortex (Krüger and Aiple, 1988). Each firing denoted
with one hidden layer can approximate any continuous, by a short vertical bar, with a separate row for each
analog function arbitrarily well. Important to many neuron. For comparison, two vertical lines mark a
implementations is the fact that second generation length of an interval of 100 msec. The time span is
networks support learning algorithms based on gradient known to suffice for the completion of some complex
descent, such as error back-propagation. multilayer cortical computations (reproduced from
Maass, 1997).
The third generation of artificial neural networks is based
on spiking neurons, or “integrate and fire” neurons. These
neurons use recent insights from neurophysiology,
often there is information inherent in the differences in
specifically the use of temporal coding to pass
timing of stimuli perceived by a biological system. In
information between neurons. These networks, like those (Thorpe et al., 2001) a number of natural situations where
of the second generation, can approximate continuous this stimuli time differential are presented, including
functions arbitrarily well, but with temporally encoded sound localization where differences in the arrival times
inputs and outputs (Maass, 1997; Maass, 1999). Further, of stimuli to the left and right ear provide a means to
there are function that require fewer neurons in a pulsed
locate the source, and motion perception in the visual
neural net to approximate than would be needed in a
system by means of timing differences between different
second generation network (Maass, 1997). photoreceptor responses.
All three of these generations are simplifications of what Many biological neural systems not only use temporal
is known about the physiology of biological neurons but information but do so with incredible precision and
the third generation is the model with the highest fidelity. accuracy. For example, the auditory system of the barn
owl has the capability to locate sources of sound in the
2.2 Biological Basis horizontal plane with a precision of 1 to 2 degrees. This
equates to a temporal difference of only a few
As was mentioned previously, one of the main impetuses microseconds (< 5µs) between the arrival of sound waves
for considering the use of temporal codes is the speed at the left and right ears (Gerstner et al., 1999).
with which they can transmit data when compared to rate
coding. Many tasks performed by biological neural This ability to learn and act in a dynamic environment,
networks are completed in a very short time, often a rich with temporal information, is a necessary quality for
period so short that only one firing per neuron could be biological systems and for artificial systems that seek to
sampled. This is demonstrated by figure 1. Note the low perform similar tasks.
number of spikes from any individual neuron over the
four second period. 2.3 Benefits of Pulsed Neural Networks
Besides the possible speed improvements, biological
Speed. As has been presented, networks composed of
systems have other reasons to use temporal codes. Very
spiking neurons can transmit and receive substantial
2
amounts of data through the relative timing of only a few
spikes. This leads to the possibility of very fast and
efficient implementations.
Real-Time Action. Pulsed networks are designed to use
temporal information and so can integrate readily into
“real” dynamic environments.
Complexity. Pulsed networks can compute any function a
second generation network can and can often do so with
fewer neurons.
Biological Fidelity. Because pulsed networks adhere more
closely to what is already known about biological neural
networks, they can benefit more readily from the rapidly
increasing base of knowledge gained from the field of
neurophysiology.

Figure 2: Potential rising past a threshold and firing. The

3 SPIKING NEURON MODEL neuron will not fire again until the refractory period,
Artificial spiking neurons are a fairly new technology, determined by _v , has ended.
being used extensively in only the last five to ten years.
Most of the publications regarding spiking neurons still
focus on research rather than implementation. Therefore a 3.2 Threshold functions
standardized model has not coalesced. Such a standard As mentioned earlier, natural neurons do not usually
may never emerge, because the category of spiking exceed a firing rate of around 100 Hz, so clearly the
neurons is already broad enough to include several neuron does not fire at all possible times when its
important subcategories. Here we describe the most potential is above the resting threshold. In nature as well
salient features common to all models and introduce some as in the spiking model, a refractory period following
of the more important variations. each firing prevents the neuron from firing for a minimum
duration, apparently around 10 msec in a real neuron. In
the spiking model, this phenomenon is mimicked by the
3.1 Behavior use of the _ function which defines the effective threshold
Let us first look at the behavior of an individual neuron potential of the neuron. As shown in figure 3, the value of
over time. An artificial spiking neuron has a value the effective threshold shoots up immediately to infinity
analagous to the membrane electrical potential in a real upon firing, thereby preventing any potential from being
neuron, which is typically referred to as its potential or above the effective threshold. After a minimum duration,
Pv(t) to indicate a potential for neuron v at a certain time t. called the absolute refractory period, the _ function
This potential is equal to the sum of its excitatory defines a period where the neuron remains reluctant to
postsynaptic potentials (referred to as EPSPs) and its fire, the relative refractory period. Figure 3 shows this
inhibitory postsynaptic potentials (IPSPs). The inhibitory function as it would be found in nature, with a slope
potentials being negative, they tend to lower the resultant delineating the relative refractory period. For the purposes
potential. If this potential becomes large enough, it may of much research, however, the approximate model on the
pass a given threshold potential, at which point the neuron right will suffice. This can simplify calculations for proof
will fire (fig. 2); thus they are often referred to as and simulation when the neurons can be assumed to fire
integrate and fire neurons. far apart in time, or in other words, the value of (t - t’) at
the next firing time will be sufficiently large that _v(t - t’)
will effectively be 0 in all cases.

3
 customary to define a length of time after which the
response function can be approximated as zero.

Common response functions include a step, or piecewise

constant, function, shown as type A in dark blue on figure
5 below; a continuous and piecewise linear function, type
B in green below; and type C, any c o n t i n u o u s l y
−4 t
( )
t − s′
decreasing function such as this one e proposed in
(a) (Jahnke et al., 1999), shown in red.

(b)

Figure 3: (a) Typical shape of a biological threshold

function. (b) Effective approximation of this function
(Maass, 1997).

3.3 Response Functions Figure 5. Response functions of increasing fidelity to

natural responses. Time 0 is equivalent to s; the point
The voltage of one spike arriving at an EPSP over time is where each graph ascends is point s’.
modelled below. It has a distinctly non-vertical leading
edge and a less steep descent thereafter, approaching 0
Each level of fidelity to the natural function shown above
volts more and more gradually.
adds more functionality to the network. Maass has proven
that some functions impossible to approximate with type
A neurons can be handled by type B (Maass, 1997). Since
the pattern here and in the advances of artificial neural
networks in general implies that fidelity to nature
improves capability, we propose the following more
refined model which more closely mimicks the natural
response function, shown in blue above:

 
 6+ 4 2 
 
Figure 4: Shape of a biological response function (Maass,
− cos  π  +1
1997).   ′ 
2
 
  t − s  + 1+ 2  
One spike is generally not enough to cause a postsynaptic    t end − t ′  π 
neuron to fire. Biological neurons typically must have  
their potentials raised by around 20 mV, while individual 2
spikes change this potential by a few millivolts at most
(Maass, 1997). Therefore to determine Pv(t) it is necessary where t - s’ represents the time since the spike reached the
to integrate the response function for each spike neuron, tend - t’ represents the duration of the applicable
encountered before time t. This could be a daunting task response function values. This equation effectively
as spikes accumulate over a simulation. Accordingly, it is stretches a cosine wave shrunk to [0, 1] onto the range of

4
all positive numbers (t - s’) using a quadratic mapping (b) The effect of several spikes arriving at one synapse
β u, v
function of the form .
(x + α )2 ∑ ε (t − s′ )
u ,v n
n:sn∈Fu

3.4 Weights
where Fu is the set of firing times of neuron u.
As in the models of the previous two generations of
artificial neurons, the input potential Pv(t) a spiking
(c) The effect of spike trains from several input neurons
neuron v are mediated by a weight wij where i is the index
of the presynaptic neuron for which this w applies. u1 Ku n
:
Generally, this is used as a multiplicand with the value of
_(t - s) to arrive at the effective EPSP or IPSP for the
connection from ui to vj as it is in the perceptron model. ∑ ∑ ε (t − s′ )
i:un∈Γv n:sn∈Fui
u ,v n
However, it has been shown that real synapses do not fire
in a deterministic manner (Maass, 1997). Rather, the
property represented by our w value would be analogous where Γv is the set of all input neurons to v.
to a probability of firing at a real synapse. The vesicles
containing neurotransmitters to cross the synaptic gap
release their contents completely, yet the release is not (d) Finally, with the effect of weights wij :
guaranteed to happen consistently. There are currently
many researchers working with these stochastic networks
which are essentially the same as networks known as the
Pv (t ) = ∑ ∑ ε (t − s′ ) w
i:un∈Γv n:sn∈Fui
u ,v n ij

spike response model. For our relatively primitive

experiments, we treated the weights as deterministic
analog values.
3.7 Information Coding
Clearly the spiking model is fundamentally different than
3.5 Excitatory vs. Inhibitory Neurons previous generations of artificial neurons. Most
In addition to the weight factor, artificial spiking neurons importantly, the information passed by spikes can only be
that of the relative timing between them. Thus the passing
can have an excitatory or inhibitory effect on the potential
of useful information across a spiking net requires
of their output connections. In the inhibitory case,
ε (t − s )
conversion from other forms (typically analog) to
is simply a negative image of the excitatory temporal data. We will see that some methods of coding
version. Recent research has indicated that neurons in the rely on relative timing alone, but that others rely on a
brain normally have one or the other effect on each synchronized, known time and input interval. While some
neuron which receive their outputs. Therefore, this of the absolute time dependent models show encouraging
variable can be modeled as a value ku which is equal to 1 results, real neurons do not have this facility and we can
assume that an effective spiking network need not include
or -1 for a particular neuron u. In practice, this k value is such absolute time sensitivity. On the other hand, it has
( )
simply assumed to be part of the ε u ,v t − s response been shown that synchronization of spike trains can be
accomplished in large populations of neurons and can
function at the synapse u, v .
facilitate learning of in some networks (Maass, 1999).

The most familiar and intuitive form of coding is known

3.6 Formal Specification for Pv (t ). as delay coding. Delay coding arises from the observation
that in the natural world, more highly stimulated neurons
Reviewing the variables described above, we can tend to spike sooner, in addition to more frequently. This
compose the formal definition of the potential value model relies on an absolute time frame, beginning at
Pv (t ). Tin and lasting for a period c. From whatever form the
input data arrived, a standard sigmoid or hyperbolic
(a) The effect of one spike traveling from u to v: tangent function must map the input onto the range [0,1].
After this transformation is made, the input x j is
ε u ,v (t − s ) converted to a time Tin − cx j when the neuron will fire.

5
 10! orders. Furthermore, the model of neuronal dynamics
which can accomplish a sensitivity to order is alluringly
simple.

Figure 6: A delay-coded neuron (Maass, 1997).

c represents the time interval for valid input and output

spikes. Input spikes outside this time interval are not
considered. Generally, the earlier the presynaptic neurons
fire, the earlier the postsynaptic one will fire. Thus the Figure 7: A simple network designed to favor the spike
first firing of the neuron is determined by the timing and of the winning neuron (Thorpe, 2000).
synergistic coincidence of the input spikes. While it is
possible that the output neuron will spike more than once,
In this model, neuron I inhibits neuron N. Therefore an
its output is determined by the first. If that time t j is accumulation of spikes which activates N will at the same
time tend to activate I, whose inhibitory factor will
Tout − t j
applied to the function yj = , y j will be a decrease the likelihood that later spikes will cause N to
c fire.
value in the range [0,1] which can then be reconverted to
the proper output data range. Other significant research into proving qualities of
various classes of spiking networks and exploring
The binary coding scheme is also based on knowledge of learning behaviors is being done at the university in Graz,
an absolute time value. In that model, spikes fill out bits Austria, by Wolfgang Maass, Berthold Ruf, and Heinrich
in a numeric value. As small time segments pass, those Schmitt. The field is still very young but there are already
segments with a spike occurring are considered to many promising avenues to explore.
represent 1; segments with no spikes represent 0.

A count code, where values are represented by the 4 LEARNING

number of spikes within a given time interval, is also
absolute time dependent.
4.1 Unsupervised Learning
The last absolute time dependent encoding method is Much of the research into learning algorithms for pulsed
called the timing code. This depends on being able to neural networks has been focused on unsupervised
determine very accurately the exact time of fire, such that learning. The aim of these algorithms is to allow the
each spike could convey a value with information network to self-organize and eventually learn to
equivalent to the inverse of the granularity of the discriminate between input patterns with no explicit
distinguishable time intervals. We did not experiment identification of those patterns. This is precisely how
with these models, as our interests are in the paradigms biological networks learn. There is no outside force
more useful for real-time response. instructing the brain as to how it should react to each
input that is presented to it. Rather, through interactions
Simon Thorpe and his team (Arnaud Delorme, Rufin internal to the brain, the outputs of the brain become
VanRullen, Marie Fabre-Thorpe) in Toulouse, France, are appropriate for the given inputs, it learns.
at the leading edge of research into spiking networks.
Among many other things, they have inspired the first Perhaps the most studied and developed is the Kohonen
commercial product, SpikeNET, a visual-recognition Self-Organizing Map (SOM) (Kohonen, 2001). Kohonen
system. They argue that a system of rank order coding, proposes that the “feature map” created by the SOM
where first spikes received by a neuron are given algorithm can be used for preprocessing patterns for
influence and later ones are inhibited, can deliver an recognition or to project and visualize high-dimensional
extraordinary amount of information (Thorpe, 2000). It is signal spaces on a two-dimensional display. Further,
immediately apparent that, as they claim, 10 input Kohonen states, “As a theoretical scheme the adaptive
neurons with even just one spike per neuron can arrive in
6
SOM processes, in a general way, may explain the methods for single layer networks include Hebb learning,
organizations found in various brain structures”. perceptron learning, and the delta rule.
Hebb learning increases weights between neurons when
both neurons are “on” at the same. This type of learning is
one of the most simple and is originally based on biology,
where this type of weight strengthening is thought to play
a major role in learning. An extended version of the Hebb
rule also strengthens weights when both neurons are “off”
at the same time. This addition makes the algorithm more
computationally powerful. The Hebb learning algorithm
is applied once to each training vector and then halts.
Training any more than this would have no effect.
The perceptron learning algorithm is more powerful than
Hebb learning. It has been proven to converge to the
correct weights if they exists for the given network
configuration and problem domain. This algorithm
Figure 8: Spikes at synapse 〈u, v〉 and potential of neuron updates the weights only when the network makes an
v during a learning cycle. In (a) the first spike of u is error in classification. In the case of an error the weights
sufficient to make v fire before the second spike of u; in are changed in proportion to a learning rate to correspond
(b) the firing of v is determined by two spikes of u. more closely to the correct answer. Perceptron learning is
(reproduced from Ruf and Schmitt, 1997) repeated on training vectors until no weights change.
The delta rule seeks to minimize the difference (error)
Very basically, the SOM is an array of nodes, each node between the real value output of the network and the
having a neighborhood of other nodes. In the learning target value. This differs from the perceptron learning
process “those nodes that are topographically close in the algorithm, which placed a threshold on the output to
array up to a certain geometric distance will activate each determine if the network responded correctly or not.
other to learn something from the same input x. This will Training using the delta rule is conducted until the error in
result in a local relaxation or smoothing effect on the the output is less than some threshold value.
weight vectors of neurons in this neighborhood, which in
continued learning leads to global ordering” (Kohonen, For multi-layer networks the most successful and studied
2001). learning algorithm, by far, is error back-propagation. This
algorithm is applied to feed-forward networks of neurons
Self-organizing maps are traditionally applied in neural which have non-linear activation functions. The error
networks of the second generation. That is, the between the output of the neurons and the target values is
information is encoded as real values representing an propagated backward through the network to adjust the
average rate of firing of a neuron. Recently, self- weights of earlier connections. This method of learning is
organizing maps, similar to those developed by Kohonen, a form of gradient descent, seeking the lowest point in the
have been implemented using networks of spiking error landscape.
neurons. Ruf and Schmitt propose a such a mechanism for
unsupervised learning in networks of spiking neurons Because Hebbian learning is biologically based, it is a
(Ruf and Schmitt, 1998). This model is a further step natural candidate for application to pulsed neural
toward biological reality in that it combines the spiking networks. Ruf and Schmitt have presented such an
neuron model with biologically plausible self- application in a model of learning based on the
organization. This implementation benefits from those differences in timing of the spikes of two neurons (Ruf
qualities of spiking neurons presented in section 2.3, most and Schmitt, 1997). By looking at the time difference
specifically speed of computation. In the self-organizing between two spikes, just one spike from both the
map, a group of neurons compete with the winning presynaptic and postsynaptic, the benefits of using spiking
neuron inhibiting the response of the other competing neurons can be maintained. Normally, learning would
neurons. With a spiking neuron the winner can be occur only after an average rate of firing was determined,
computed quickly and locally, based only on single firing as in the learning methods described above for the first
events. two generations of neural networks.
The Hebbian-type learning rule proposed by Ruf and
4.2 Supervised Learning Schmitt to change the weight w u,v from a presynaptic
neuron u and postsynaptic neuron v is
Supervised learning has a long and very well studied
history in artificial neural networks. Early training ∆wu,v = η(tv – t0)

7
where tv is the firing time of the postsynaptic neuron and is increased above 1 or decreased below 0 it is set to 1 or
t0 is the second firing time of the presynaptic neuron (tu is 0 respectively. Over time, this rule leads to neuron v firing
the first firing time of neuron u), and η is the learning at desired times and remaining inactive at other times.
rate. If the second firing of u, t0, occurs before the firing
Because training in this way only occurs when the
of v (tv – t0 > 0) then the weight between them, w u,v ,
postsynaptic neuron fires it is crucial that the neuron
should be increased. If the opposite occurs, tv is before t0,
always has a chance to fire, even when the initial weights
the weight should be decreased. This is shown by figure
for its synapses are very low. There are two ways to
8, which compares these two cases.
ensure that a neuron always has at least some chance of
The assumption is that the first firing of the presynaptic firing. The first is to have a dynamic threshold. When the
neuron, u, is intended to make the postsynaptic neuron, v, neuron fires the threshold is increased by some value.
fire at time t0, along with the second firing of u. So, the Over periods of inactivity the threshold is decreased. So,
firings of the presynaptic and postsynaptic neurons are to if a neuron does not fire for an extended period of time
be synchronized and the network will have learned to the threshold will lower enough such that it begins firing
output the desired spike train when given a specific input. and hence learning. The second method is to have a base
firing rate that the neuron maintains even when it receives
The above method includes only one presynaptic and one no activation. In this case also the neuron has a chance to
postsynaptic neuron. In most situations, however, a learn periodically even though it may not be receiving
postsynaptic neuron will have many presynaptic neurons enough activation due to low weights. This base firing
contributing to its activation. In this case the weights rate exists is known to exist in biological neurons. In this
could still be learned in the same manner, but the case we chose to implement the first method, a dynamic
postsynaptic neuron could only receive input from only threshold.
one presynaptic neuron at a time. The weights in this case
would be learned serially.
The weights can also be learned in parallel through an
expansion of the above method. The postsynaptic neuron
in this case is induced to fire at some time t0 by way of an
additional synapse with sufficient weight. The weights
wui,v on the synapse from neuron ui to neuron v are
updated by the following rule
∆wui,v = η(t0 – tui)
where t ui is the firing time of neuron ui. Further, the
weight vector must be normalized after each application
of the learning rule.

5 IMPLEMENTATION OF LEARNING
Our implementation of supervised learning with a spiking
neuron, like that presented by Ruf and Schmitt, is based
on Hebbian learning. That is, synaptic weights are
increased when both the presynaptic neuron is active at
the time that the postsynaptic neuron produces a spike. Figure 9: Three types of rhythms used in training. a)
The weights in this case are initialized randomly in the The top shows a high frequency spike train on the first
range from 0 to 1. The rule for the change of weight w ui,v input and a low frequency spike train on the second
from neuron ui to neuron v whenever neuron v produces a input. Below this is the opposite. Correspondingly, the
spike is first output neuron should spike for the top set and the
second output neuron should spike for the bottom set.
∆wui,v = η εu(t – tu) χ b) A training set where the inputs are either
where η is the learning rate, εu(t – tu ) is the response of synchronized or not. c) A training set where either the
neuron u at the time, t, that neuron v fires, and χ is an first input regularly occurs shortly before or shortly
indicator of whether the spiking of neuron v is desired at after the second.
this time. If the spiking is not desired, χ is –1, otherwise it
is 1. In other words, the weight of the synapse from u to v A set of simple test data was constructed to evaluate the
will change by an amount proportional to the contribution learning rule above. Earlier generations of neurons, based
of neuron u at the time of firing and will be positive if the on rate coding, often use simple functions of two inputs to
spiking is desired and negative if not desired. If a weight test neural models and learning, such as the logical AND,

8
Figure 10: Top: Action potential of two output Figure 11: Top: Action potential of two output
neurons while training on frequency rhythms. A “1” at neurons while testing after training. Bottom: Spikes
the top of the graph signifies that a high frequency generated by output neurons. The question marks
input is being given on input one and that output one signify that the network does not know which output it
(blue) should be more active. A “2” signifies that the is expected to provide.
opposite pattern is being presented and that the second
output (green) should be more active. The dashed lines
represent the dynamic thresholds of the output neurons. 5.2 Results
Bottom: The corresponding spikes of output one and
two. A vertical line is one spike, blue for 1 and green The graphs in figure 10 show the results of training a
for 2. network of spiking neurons with two inputs and two
outputs using the frequency based rhythms as training
vectors. In the case of a high frequency spike train on the
OR and XOR functions. We sought to create equally first input and a low frequency spike train on the second
simple functions based on temporal coding. input the first output neuron should be active, that is, have
a higher action potential and spike more often. In the case
5.1 Rhythms of the reverse the second neuron should be the more
active.
We developed three “rhythms” to test the presented
supervised learning algorithm for spiking neurons. The You can see that after one a few presentations of data the
networks in this case are composed of two input neurons, appropriate neurons begin to fire when presented with the
which receive the rhythmic sequences of firing, and two frequency patterns. While encouraging at first this result
output neurons, that will fire to signal which rhythm is is actually misleading, as is shown by figure 11. These
being presented. The three rhythms are based on the graphs show the response of the network after it has been
relative frequency of spikes, the synchrony of spikes, and trained and the weights are stationary. The network was
the relative timing of spikes. These rhythms are presented presented with the same frequency rhythm data that it was
graphically in figure 9. trained on. You can see that one neuron is continually

9
active while the other hardly responds at all. The
promising results during training were actually created by The networks that can be created on the simulator are
the changing of the weights and not the weights highly configurable. Individual neurons may be added at
themselves. Each time a neuron is supposed to respond its any time and connected by dragging the mouse from the
weights are increased, hence making it respond more. source neuron to the destination. New layers can be added
Correspondingly, when it is desired that a neuron not fire en masse, with the options partially shown in this box:
its weights are decreased. This occurs each time that a
pattern is presented. The network is small enough such
that it can respond to these changes rapidly enough that it
produces the correct responses for most of the time.
Because the second neuron (graphed as green) was the
last one to be trained positively during the learning
sessions, it is the one that responds to the inputs in the
testing set.

6 SPIKING NETWORK SIMULATOR

In order to experiment with and observe the dynamics of
spiking nets, we created a Windows application which
facilitates the creation of a pulsed neural network and
provides rudimentary input and output resources. A
screen shot is given below. Each layer can be connected to any of the previous layers
where each neuron is connected to either all, none, or only
those neurons which are in a range of the relative position
of the destination neuron in its layer. Competing
connections (connections within the same layer) can also
be specified in the same manner. All competing
connections are made inhibitory by default. This would
allow us to experiment with Kohonen map-style self-
organizing learning.

While training, neurons in the display window show their

relative potentials as a shade of yellow and flash red when
they fire. The coincidences of rhythm between neurons
and the general activity level of each neuron is thus
immediately visible.

The simulation shown above was from a layer of seven Each of the dynamics of our learning model were
input neurons to two outputs. The potentials of the two configurable by the control panel on the right. The
following shows our default experimental starting points:
output neurons are recorded by the scrolling graph at the
upper right.

Each input neuron may be fired by pressing an

alphabetical key, where the letter “a” is mapped to the
first neuron in the layer, “b” to the second, and so on.
This was the chosen input method because (1) there
needed to be a sensible way to interact with the input
layer in real time, and (2) we were curious as to the
capabilities of neural nets and especially spiking nets to
recognize users by the rhythm of their typing.

Target outputs are chosen by pressing a numeric key “1”

through the number of neurons in the output layer. This
target was then trained while the simulator ran by our
pseudo-Hebbian algorithm which will be described
shortly.

10
The default threshold is given as a multiple of the 2 Gerstner, W., Kempter, R., Leo van Hemmen, J.,
maximum response function voltage. The threshold boost Wagner, H. (1999) “Hebbian Learning of Pulse
is a factor applied to neurons when they fire, partly to Timing in the Barn Own Auditory System”. In W.
make up for their eventual loss of resistance through the Maass and C. M. Bishop (eds.), Pulsed Neural
periodic threshold decay, but partly to reduce their Networks. MIT Press, Cambridge, Mass., 1999.
sensitivity when overstimulated. It is our observation that 3 Jahnke, A., Roth U., Schönauer, T. (1999) “Digital
biological neurons consistently find a middle ground Simulation of Spiking Neural Networks”. In W.
between not firing at all and firing at their maximum rate. Maass and C. M. Bishop (eds.), Pulsed Neural
Since neurons, like all cells, are essentially autonomous Networks. MIT Press, Cambridge, Mass., 1999.
entities within the protection of the environment held
constant by the body around them, it seems sensible that 4 Kohonen, T. (2001) Self-Organizing Maps. Springer
Series in Information Sciences, Vol. 30, Springer,
their gravitation toward the middle ground of firing (30-
Berlin, Heidelberg, New York, 2001.
50Hz) is critical to their health. The learning rate is the
η found in our learning equation, 5 Krüger, J., Aiple, F. (1988) "Multielectrode
investigation of monkey striate cortex: spike train
wnew = wold + η εu(t – tu) χ correlations in the infragranular layers", J.
Neurophysiology, vol. 60, pp 798--828, 1988.
where t is the target value and ε represents the strength of 6 Maass, W. (1997) “Networks of Spiking Neurons:
the EPSP at the time of firing. The Third Generation of Neural Network Models”
Neural Networks, 10(9):1659–1671.
The simulator was used to verify the MATLAB results as
7 Maass, W. (1999) “Computing with spiking
per learning rhythms as well as to get immediate results
neurons”. In W. Maass and C. M. Bishop (eds.),
with various experiments. Many configurations of
Pulsed Neural Networks. MIT Press, Cambridge,
network were tried and trained within a short period to Mass., 1999.
gauge their abilities to learn. Though our nets did not
realize a satisfying learning achievement, we ourselves 8 Ruf, B., Schmitt, M. (1997) "Learning temporally
learned to recognize many unpredicted dynamics of encoded patterns in networks of spiking neurons,"
spiking nets through the observations of our data. The Neural Processing Letters, vol. 5, no. 1, pp. 9--18,
program is available at 1997.
9 Ruf, B., Schmitt, M. (1998) "Self-Organization of
http://www.cs.rit.edu/~ccm2451 /pub/snn.exe Spiking Neurons Using Action Potential Timing", in
IEEE Trans on Neural Networks, Vol. 9, No. 3, May
for anyone who is interested in using it. 1998, pp. 575-578.
10 Thorpe, S., Delorme, A., VanRullen, R. (2001)
“Spike based strategies for rapid processing”. Neural
7 REFERENCES Networks, 14(6-7), 715-726.
1 Gerstner, W. (1999) “Spiking Neurons”. In W. Maass
and C. M. Bishop (eds.), Pulsed Neural Networks.
MIT Press, Cambridge, Mass., 1999.

11