=Paper=
{{Paper
|id=Vol-1525/paper-21
|storemode=property
|title=Pattern recognition with spiking neural networks: a simple training method
|pdfUrl=https://ceur-ws.org/Vol-1525/paper-21.pdf
|volume=Vol-1525
|dblpUrl=https://dblp.org/rec/conf/splst/ChristopheMAKL15
}}
==Pattern recognition with spiking neural networks: a simple training method==
https://ceur-ws.org/Vol-1525/paper-21.pdf
SPLST'15
Pattern recognition with Spiking Neural
Networks: a simple training method
François Christophe, Tommi Mikkonen, Vafa Andalibi, Kai Koskimies, and
Teemu Laukkarinen
Tampere University of Technology
Korkeakoulunkatu 1, FI-33720 Tampere, Finland
firstname.lastname@tut.fi
Abstract. As computers are getting more pervasive, software becomes
transferable to different types of hardware and, at the extreme, being
bio-compatible. Recent efforts in Artificial Intelligence propose that soft-
ware can be trained and taught instead of “hard-coded” sequences. This
paper addresses the learnability of software in the context of platforms
integrating biological components. A method for training Spiking Neu-
ral Networks (SNNs) for pattern recognition is proposed, based on spike
timing dependent plasticity (STDP) of connections. STDP corresponds
to the way connections between neurons change according to the spiking
activity in the network, and we use STDP to stimulate outputs of the
network shortly after feeding it with a pattern as input, thus creating
specific pathways in the network. The computational model used to test
this method through simulations is developed to fit the behaviour of bi-
ological neural networks, showing the potential for training neural cells
into biological processors.
Keywords: Pattern recognition, Artificial Neural Networks, Spiking Neu-
ral Networks, Computational models, Computational Biology
1 Introduction
Software is everywhere: the human environment is populated by more and more
software-driven intelligent devices, connected by Internet and other networks.
With the apparition of wearables and implantables, computers are getting more
and more pervasive and close to the biological world. In such systems, software
is expected to expand on various types of platforms.
In the current information technology, the interplay between biology and
software has been indirect. Humans use software through various user interfaces,
rather than with direct communication links. Concepts from biological systems
have inspired various heuristic algorithms to solve computer science problems
(typically optimization and search), or inspired software for communication sys-
tems to mimic the adaptive behavior of biological systems [15, 12]. Novel ways
of programming by training, teaching, imitation and reward are already being
296
SPLST'15
demonstrated in robotics with the help of in-silico chips behaving like neurons,
i.e. neuromorphic chips [7].
The work reported in this paper is a first step in a project aiming at devel-
oping techniques to support the direct run-time interaction of biological entities
and software. Our vision is that eventually software interacts directly with the bi-
ological world: software controls biological entities, and biological entities control
software systems. Thus, rather than using the biological world as a model of new
algorithms, we intend to let biological entities communicate directly with soft-
ware. In this way, software systems and biological entities form co-operational or-
ganizations in which both parties solve problems suitable for them, contributing
to a common goal. Typically, biological entities are superior in efficient massive
parallel processing of fuzzy data and resilient to damage, while traditional soft-
ware systems are better suited for making discrete, well-defined logical decisions.
Biological entities are also far more energy-efficient than traditional computing
devices.
The project focuses on integrating real neural cultures with software systems.
A central problem then is the training of biological neural structures for a partic-
ular task and the connection of the trained neural culture to the software system.
However, to experiment with different approaches to solve these problems, avail-
able biological neural cultures impose many practical problems: their detailed
structure is difficult to study, their lifetime is limited, and they need constant
nutrition. Luckily, for the past decades, Artificial Neural Networks (ANNs) have
evolved to the point of being currently very close in behaviour to biological neural
structures [13]. Thus, in the first stage of the project, we use ANNs to simulate
biological neural networks. In a later stage we aim to transfer the techniques to
biological neural cultures currently available on Multi-Electrode Arrays (MEAs)
[16].
In this paper we study the basic training problem of biological neural net-
works using a biologically realistic model of spiking neurons. A simple pattern
recognition problem is applied to this model. We demonstrate that a training
technique based on Spike-Timing-Dependent-Plasticity (STDP) appears to be
sufficient for these kinds of tasks.
The rest of this paper is structured as follows. In Section 2, we discuss related
work. In Section 3, we introduce the computational model used in this paper. In
Section 4, we evaluate the results we have obtained. In Section 5, we draw some
final conclusions.
2 Related work
In [11], Maass draws a retrospective of the techniques used for modeling neural
networks and presents the third generation of neural networks: SNNs. This study
297
SPLST'15
classifies neural networks according to their computational units into three gen-
eration: the first generation being perceptrons based on McCulloch-Pitts neurons
[4], the second generation being networks such as feedforward networks where
neurons apply an “activation function”, and the third generation being networks
where neurons use spikes to encode information. From this retrospective, Maass
presents the computational advantages of SNNs according to the computation
of, first, boolean functions, and secondly according to functions with analog in-
put and boolean output. For instance, Seung demonstrated in [14] that SNNs
can be trained to behave as an XOR logical gate. The study of single neurons,
the population of networks and plasticity [3] provides guidelines on how single
computation units, i.e. neurons, function but more importantly on how to struc-
ture a network (e.g. number of layers, number of units in a layer) and on the
models of evolution of connectivity between neurons, i.e. plasticity.
In their chapter on computing with SNNs [13], Paugam and Bohte present
different methods applied for learning in SNNs. In this review chapter, they
distinguish the traditional learning methods issued from previous research with
ANNs, and learning methods that are emerging solely from computing with
SNNs. Among the traditional methods are temporal coding, unsupervised learn-
ing such as Hebbien learning or Kohonen’s self-organizing maps, and supervised
learning such as error-backpropagation rules. About the “unconventional” learn-
ing methods, they are regrouped into so-called Reservoir Computing methods.
Reservoir Computing methods regroup Echo State Networks and Liquid State
Machines. The main characteristic of reservoir computing models relies in the
apparent disorganization of the network between input and output layers. This
network, the reservoir, is a recurrent network where neurons are interconnected
by a random sparse set of weighted links. More over, this network is usually left
untrained and only the output connections are trained and optimized according
to the desired answer based on what input is given.
To a certain extent, the simple training method proposed in this paper follows
the idea of training only the output layer as STDP will act alone for reinforcing
the positive pathways of the network. However, the network studied in this paper
is simply a feed-forward network modeled with the BRIAN simulator [2].
3 Computational model
The training method presented in this paper is studied with the computational
model presented in this section because it provides a first test of feasibility before
testing this method on biological Neural Networks (bioNNs). The model used for
this research is composed of two basic elements: neurons and synapses. Neurons
are built based on the firing model of Izhikevich which is shown to be very
realistic in [9]. Synapses follow the Spike Timing Dependent Plasticity (STDP),
meaning that a connection between two neurons will grow if the post-synaptic
neuron fires soon after the pre-synaptic neuron. On the opposite, a connection
298
SPLST'15
will decrease if the post-synaptic neuron fires before the pre-synaptic neuron.
This section presents these two basic models in more details and then gives a
view on the composition of the entire neural network.
3.1 Model of spiking neuron
The model of a spiking neuron used in this study is from Izhikevich [8]. This
model reproduces the dynamic behavior of neurons while being computationally
simple as opposed to models accounting for the structural parameters of neurons
(for example, Hodgkin-Huxley model [6]). The Izhikevich model expresses the
variations of electric potential in the neuron’s membrane according to the current
flowing through the membranes ion channels. These electric potential variations
are expressed in the form of two differential equations, as follows:
dv
C = k(v − vr )(v − vt ) − u + I
dt (1)
du
= a(bv − u)
dt
where v represents the membrane potential, u the recovery current of the mem-
brane, I the input current through the membrane, C the membrane capacitance,
vr the resting potential of the membrane, vt the threshold potential for the mem-
brane to fire a spike, k, a and b parameters adjusted according to the firing
pattern required. The variables v and u of equation 1 are reset after v reaches a
peak value vpeak , as follows:
(
v←c
if v ≥ vpeak , then (2)
u←u+d
Figure 1 presents an example of simulation of a pyramidal neuron exhibiting
regular spiking as stimulated with a constant input current of 70pA. The same
simulation among others are compared with recordings from real neurons in [9]
showing the precision of this models in reproducing these dynamic patterns while
being computationally simple.
3.2 Model of STDP
STDP is a rule for neurons to strengthen or weaken their connections according
to their degree of synchronous firing [1]. This rule mostly known in Neurobiology
and Neuroscience is similar to the Hebbian learning rule widely used in learning
Artificial Neural Networks and Self-Optimizing Maps [5, 10]. Considering a pre-
synaptic neuron i and a post-synaptic neuron j, the STDP rule characterizes
the changes in synaptic strength as:
N X
X N
∆wj = W (tlj − tki ) (3)
k=1 l=1
299
SPLST'15
Fig. 1. Example of Regular Spiking pyramidal neuron simulated with Izhikevich model
from Equations 1 and 2 (step input stimulation current I = 70pA from 0 to 1s).
Simulated with the following values of parameters: vr = −60mV, vt = −40mV, vpeak =
35mV, C = 100pF, k = 0.7pA/mV2 , a = 30Hz, b = −2nS, c = −50mV, d = 100pA.
300
SPLST'15
with the function W (x) defining the order of decrease or increase of strength
depending on the synchrony of spiking between pre- and post-synaptic neurons,
expressed as:
(
A+ exp(− τx+ ) if x > 0
W (x) = (4)
A− exp( τx− ) otherwise.
In equations 3 and 4, tlj represents the lth spiking time of neuron j; similarly,
ti stands for the k th spike timing of neuron i; A+ and A− are constants defining
k
the amplitude of change in weight (at t = 0+ and t = 0− , respectively); and, τ+
and τ− are time constants of the exponential decrease in weight change.
Fig. 2. Model of STDP (reproduced after Bi and Poo [1])
301
SPLST'15
Figure 2 represents equation 4 of relative weight changes in relation with the
experiments realized by Bi and Poo in [1]. This figure shows typically a decrease
in synaptic weight when the pre-synaptic neuron spikes after the post-synaptic
neuron, and on the opposite an increase in connectivity from pre- to post-
synaptic neuron if the pre-synaptic neuron spikes just before the post-synaptic
neuron.
3.3 Network model
The network developed in this study as an example of pattern recognition is
presented in Fig. 3. This network is dedicated at recognizing patterns from a
5x5 pixels image. It is composed of:
– 25 input neurons corresponding to each pixel of the image,
– a hidden layer of 5 neurons, and
– 2 output neurons (1 corresponding to the neuron reacting when a circle
appears in the image, the other one being a test neuron for comparison
between learning and no stimulation).
Fig. 3. Representation of the network developed for 1 pattern recognition
302
SPLST'15
4 Training method and Evaluation
This section goes more into the details of the simple learning method using
STDP rule for training the network. A pattern recognition task is used as a case
study for testing the learning method proposed.
4.1 Pattern recognition task
The pattern recognition task evaluated during this simple case study consists in
making the difference between a circle given as input stimuli and other inputs
(in that case an X-cross shape). These shapes are represented as coming from
a 25 pixels image (5x5 matrix), each pixel being binary: black or white. Figure
4 presents the input circle and X-cross patterns and their respective representa-
tions as input stimuli to the network.
(a) Circle pattern (b) X-cross pattern
Fig. 4. Input patterns as images and their respective representations as input stimuli
4.2 Training method and test protocol
After initial tests on the learning ability of the network, the training period was
adjusted to be 15s during which the input layer is stimulated every 100ms with
a circle pattern. 10ms after stimulating the input layer, the output neuron is
given an external stimulation making it to spike. This spiking, in relation with
the preliminary spiking of neurons from the input layer, reinforces the paths be-
tween activated neurons of the input layer and the trained neuron of the output
layer. This training can be seen in the first phase of the time diagrams (top and
middle) of Figure 5 from t = 0 to 15s.
303
SPLST'15
Fig. 5. Training phase and test phase of the circle recognition experiment
The testing phase is composed of 6 stimuli with circle pattern and 7 stimuli
with a different pattern (in that case representing an X-cross). These stimuli
of the input layer happen without any external stimulation of the output layer.
The neuron trained for recognizing a circle fires on its own after the learning
phase. These test patterns are sent between t = 15, 5 to 25s with the following
sequence: {circle, circle, cross, cross, circle, cross, cross, cross, cross, circle, circle,
cross, circle} at the respective times {15.5, 16, 16.5, 17, 17.5, 18, 18.5, 19, 20,
21, 23, 24, 24.5} seconds. The two upper time diagrams of Figure 5 show this
test phase.
The third time diagram of Figure 5 (down) presents the evolution of strength
of synapses between neurons from the hidden layer and neurons from the output
layer. This evolution shows first a stabilization period from the random strengths
given as initial condition to lower values. Secondly, learning can be seen as the
strengths of certain synapses increase to high levels of connectivity (i.e. to lev-
els higher than 0.8 times the maximum connectivity and often reaching this
maximum).
4.3 Results
The 1000 simulations of this experiment revealed a success rate of 80.3% in
recognizing a circle from a cross. This rate was computed after the execution
of a thousand experiments. From these experiments, five different cases were
observed:
– correct learning: output fires only when a circle is given as input (80.3%)
– some mistakes: output fires sometimes when input is a cross (5.7%)
– always firing: output always fires whatever the input may be (12.8%)
– no learning: output never fires (1.1%)
304
SPLST'15
– wrong learning: output fires only when input is a cross (0.1%)
These different cases are represented in Figure 6.
These results show a high learning rate, i.e. a high rate of correct experiments
(80.3%), meaning that such learning method has correct grounds. Indeed, there
is space for improving this rate and a lot to learn from the analysis of failed
experiments.
First, we can notice that the network keeps on learning even after the train-
ing phase has stopped. Each time a pattern is fed as input to the network, small
increase in synaptic weights take place.
The second important thing noticed due to this continuous learning is that
the output neuron trained to recognize a circle also gets trained when receiving
another pattern as input. As the synaptic levels may already be high, it requires
only few X-cross stimulation signals for the output neuron to start spiking and
we notice that when it has learned to spike for a cross pattern, it will then fire
each time a cross appears. This is what happens for the cases where some mis-
takes are found (57 cases out of 1000 simulations).
Third, for 128 simulations the synaptic levels are high from the beginning of
the training due to the initial random value of synaptic weights. This causes the
network to be “over-trained” and thus to fire for every kind of patterns from the
beginning of the test phase. In the contrary, for 11 simulations the output neuron
does not fire at all when given any input. These 11 tests show clearly low rates in
synaptic weights and low increase in synaptic weights during training expressing
the fact that the network does not have time to learn during this period of time.
The only simulation where the output neuron fires only and always when given
the wrong input could not be explained.
5 Discussion and future work
We have studied a central problem in using biological neural networks as com-
puting resources: how to train the neural culture for a particular task. The SNN
presented in this paper shows reasonable success rate in learning (80.3%) to
differentiate between two patterns. As the behavior of the simulated network is
very close to that of real biological neural networks (bioNNs), this experiment
gives preliminary insight for using bioNNs to process complex tasks requiring
massive parallelism.
However, there is still possibility for improvement in performing such recog-
nition tasks. In the case where pathways did not have time to form during initial
training, it is indeed possible to continue training the network until synaptic level
reaches the appropriate levels. On the opposite case, when the network fires for
any type of input pattern, meaning that it is “over-trained”, training the net-
work with negative output stimulation should help restoring the differentiation
305
SPLST'15
(a) Correct learning (rate 80.3%)
(b) Some mistakes (rate 5.7%) (c) Always firing (rate 12.8%)
(d) No learning (rate 1.1%) (e) Wrong learning (rate 0.1%)
Fig. 6. Example of the 5 different cases found in experiments
306
SPLST'15
between input patterns. Such negative stimulation can be realized by stimulat-
ing the output neuron just prior to the input pattern, when an input pattern
is to be discarded by the network. This way, the synaptic pathways related to
this input pattern would decrease due to the STDP rule, thus de-correlating the
output neuron with this input pattern.
The proximity in behavior of SNNs from bioNNs should not require efforts
to transfer computations from silicon platforms into biological platforms. From
this point, transferring the various tasks developed for the past sixty years with
ANNs (e.g. classifiers, generators, etc.) to small bio-computers will be possible.
Research directions for such transfer lead to the following questions:
– Can a taxonomy of the various tasks performed with ANNs and their hier-
archical relations be developed?
– Can we classify tasks as unitary/atomic to some higher level? On the other
hand, can tasks be broken down into summation of unitary tasks?
– Is it possible to automatically associate tasks to network structures (number
of nodes, number of layers etc.) ?
– Can training also be automatically generated for these specific tasks?
These questions could lead to the construction of a compiler deriving the
number of biological neural networks used for an application, their structural
parameters and the training associated to each task required for the application.
6 Conclusion
In this paper, we presented a simple learning method using STDP for training
pathways of a SNN. This method was tested on a SNN model trained to differ-
entiate between two patterns given as input to the network. The results of this
test (80.3% success rate) combined with the fact that the behaviour of the simu-
lated network is very close to the one of a real biological network gives promising
expectations for the future of this project. This first test is still a preliminary
test towards applying bioNNs to the computation of more complex tasks such
as handwritten digit and character recognition 1 . Expectations on the results to
such test should give similar success rate as to the test conducted in this paper.
Next experiments will be conducted on real biological cells in order to validate
the possibility of training bioNNs for pattern recognition tasks. In the final stage
we intend to connect such trained bioNNs with software applications requiring
pattern recognition capability such as classification of moving objects.
Acknowledgement
This research is funded by the Academy of Finland under project named “Bio-
integrated Software Development for Adaptive Sensor Networks”, project num-
ber 278882.
1
Such training is usually tested on the MNIST dataset available from
http://yann.lecun.com/exdb/mnist/
307
SPLST'15
References
1. Bi, G.Q., Poo, M.M.: Synaptic modifications in cultured hippocampal neurons:
dependence on spike timing, synaptic strength, and postsynaptic cell type. The
Journal of neuroscience : the official journal of the Society for Neuroscience 18,
10464–10472 (1998)
2. Brette, R., Rudolph, M., Carnevale, T., Hines, M., Beeman, D., Bower, J.M., Dies-
mann, M., Morrison, A., Goodman, P.H., Harris, F.C., Zirpe, M., Natschläger, T.,
Pecevski, D., Ermentrout, B., Djurfeldt, M., Lansner, A., Rochel, O., Vieville, T.,
Muller, E., Davison, A.P., El Boustani, S., Destexhe, A.: Simulation of networks
of spiking neurons: A review of tools and strategies (2007)
3. Gerstner, W., Kistler, W.: Spiking neuron models: Single neurons, populations,
plasticity. Cambridge University Press (2002)
4. Hayman, S.: The mcculloch-pitts model. IJCNN’99. International Joint Conference
on Neural Networks. Proceedings (Cat. No.99CH36339) 6 (1999)
5. Hebb, D.O.: The Organization of Behaviour: A neuropsychological theory. Wiley
(1949)
6. Hodgkin, A.L., Huxley, A.F.: A quantitative description of membrane current and
its application to conduction and excitation in nerve. Journal of Physiology pp.
500–544 (1952)
7. Indiveri, G., Linares-Barranco, B., Hamilton, T.J., van Schaik, A., Etienne-
Cummings, R., Delbruck, T., Liu, S.C., Dudek, P., Häfliger, P., Renaud, S.,
Schemmel, J., Cauwenberghs, G., Arthur, J., Hynna, K., Folowosele, F., Saighi,
S., Serrano-Gotarredona, T., Wijekoon, J., Wang, Y., Boahen, K.: Neuromorphic
silicon neuron circuits. Frontiers in neuroscience 5(May), 73 (1 2011)
8. Izhikevich, E.M.: Simple model of spiking neurons. IEEE transactions on neural
networks / a publication of the IEEE Neural Networks Council 14(6), 1569–72 (1
2003)
9. Izhikevich, E.M.: Dynamical Systems in Neuroscience : The Geometry of Excitabil-
ity and Bursting, Chapter 8. The MIT Press (2007)
10. Kohonen, T.: The self-organizing map 21(May), 1–6 (1998)
11. Maass, W.: Networks of spiking neurons: The third generation of neural network
models. Neural Networks 10(9), 1659–1671 (1997)
12. Nakano, T., Suda, T.: Self-organizing network services 16(5), 1269–1278 (2005)
13. Paugam-Moisy, H., Bohte, S.: Computing with spiking neuron networks. In: Hand-
book of Natural Computing, pp. 335–376 (2012)
14. Seung, H.S.: Learning in spiking neural networks by reinforcement of stochastic
synaptic transmission. Neuron 40(6), 1063–1073 (2003)
15. Suzuki, J., Suda, T.: A middleware platform for a biologically inspired network
architecture supporting autonomous 23(2), 249–260 (2005)
16. Taketani, M., Baudry, M. (eds.): Advances in Network Electrophysiology: Using
Multi-electrode Arrays. Springer (2006)
308