LEARNING IN HUMAN NEURAL NETWORKS

sciencediscussionAI and Robotics

Oct 20, 2013 (3 years and 1 month ago)

147 views


1

LEARNING IN HUMAN NEURAL NETWORKS

ON MICROELECTRODE ARRAYS


R. Pizzi*, F. Gelain°, D. Rossetti* & A. Vescovi°

*Department of Information Technologies, University of Milano (Italy)

via Bramante 65


26013 Crema (CR)


tel. 39 02 50330072, fax 39 02 503300
10

e
-
mail
pizzi@dti.unimi.it

° Stem Cells Research Institute, DIBIT S.Raffaele

via Olgettina 58


20132 Milano (Italy)


tel. 39 02 21560202

e
-
mail
gelain.fabrizio@hs
r.it





ABSTRACT


Researchers of the Department of Information Technologies of the University of Milano and of
the Stem Cells Research Institute of the DIBIT
-

S. Raffaele Milano are experimenting the growth
of human neural networks of stem cells on a MEA

(Microelectrode Array) support.

The Microelectrode arrays (MEAs) are constituted by a glass support where a set of tungsten
electrodes are inserted.

We connected the microelectrodes following the architecture of classical Artificial Neural
Networks, in pa
rticular Kohonen and Hopfield networks. The neurons are stimulated following
digital patterns and the output signals are analysed to evaluate the possibility of organized
reactions by the natural neurons.


2

The neurons reply selectively to different patterns

and show similar reactions in front of the
presentation of identical or similar patterns.

Analyses performed with a special Artificial Neural Network called ITSOM show the possibility
of codifying the neural responses to different patterns.

A new experim
ent with more complex patterns has been carried out and analyses of the results are
underway.

We aim to design further experiments that improve the hybrid neural networks capabilities and to
test the possibility of utilizing the organized answers of the ne
urons in several ways.


Keywords:
neural networks, stem cells, microelectrode arrays, learning, self
-
organization.





1. INTRODUCTION


During the past decade several laboratories in the world have carried out experiments on direct
interfacing between elec
tronics and biological neurons, in order to support neurophysiological
research but also to pioneer future hybrid human
-
electronic devices, bioelectronic prostheses,
bionic robotics and biological computation.

As microelectrodes implanted into brain give r
ise to rejection and infections, researches are under
way, experimenting a direct adhesion between electronics and neural tissue, achieving important
results [13, 1, 63, 7, 5, 62, 4, 8, 32, 56, 39, 34, 49, 58, 6 , 33, 48, 66, 2 and 3].


During the early ni
neties a direct interface between nervous cells and silicon has been established.
In particular leech neuron have been used, because of their big size [22]. The Fromherz’s group

3

(Max Planck Institute of Biochemistry) first pioneered the silicon/neuron inte
rface and keeps
developing sophisticated techniques to optimise this kind of junction [18, 19, 20,21 and 17].

Many other experiments have been carried out, with different aims: in 1999 William Ditto and
collaborators at the Georgia Tech tried to obtain si
mple computations from a hybrid leech
-
electronics creature. As the neurons don’t behave as “on
-
off” elements, it has been necessary to
send them signals and interpret the neural output using the chaos theory [57, 37, 23].

In 2000 a team of the Northwest
ern University of Chicago, University of Illinois and University of
Genoa [49] developed a hybrid creature consisting of lamprey neurons connected to a robot. In
front of light stimuli, the creature behaves in different ways: follows light, avoids it, mov
es in
circle.

In 2002 Steve Potter (Georgia Tech) created a hybrid creature made by
few thousand living
neurons from rat cortex placed on a special glass Petri dish instrumented with an array of 60
micro
-
electrodes
, also able to learn from environment [1
2].

In 2003 the Duke University’s group [9] succeeded in connecting 320 microelectrodes to
monkey cells in the brain, allowing to directly translate the electrical signals into computer
instructions, able to move a robotic arm. This will be the way to a
llow disabled people to move
paralyzed limbs or electronic prostheses.


Despite of these astonishing results, the neurophysiological research is far from understanding in
detail the learning mechanism of the brain and fails to interpret the cognitive mean
ing of the
signals coming from the neurons. A deeper and innovative analysis of the signals coming from a
direct connection between electronics and neural tissue could disclose new prospects in this field .
On the other hand, a correct interpretation of th
e signals could support a faster development of the
above described bionic techniques.



4

With this purpose our group, formed by researchers of the Department of Information
Technologies of the University of Milano and of the Stem Cells Research Institute of

the DIBIT
-
S. Raffaele Milano, is experimenting the growth of human neural networks of stem cells on a
MEA (Microelectrode Array) support.


In order to examine the neural learning and memorizing activities, we developed architectures
based on the Artifici
al Neural Network models on networks of human neural stem cells adhering
to microelectrode arrays (MEAs).

The MEAs are connected to a PC via an acquisition device that allows to stimulate the neurons
with suitable inputs and to acquire the neuron signals
in order to evaluate their reactions.

In this paper we will describe the techniques we used to perform this task.

By stimulating cells with digital bitmap patterns, we recorded their electrical reactions and
examined the output signals with advanced techni
ques. This allowed us to evaluate the self
-
organization increase of the biological neural networks during and after the learning, and to
discriminate their reactions to different patterns.




2. MATERIALS AND METHODS


2.1 The stem cells


Our neurons have

been cultured starting from human neural stem cells extracted by a human
embryo. Stem cells are multi
-
potential undifferentiated cells whose main features are the ability of
self
-
renewal and of differentiation into several types of adult cells [61, 24].


5

W
e chose to adopt human stem cells because their well
-
known

capability to integrate with the host
tissue in transplantation procedures. This could allow in the future a direct implantation of a
bionic chip into a human neural tissue without rejection.


More
over, stem cells have the advantage to develop a cellular line , thus a virtually infinite
number of standardized cells, whereas the use of other types of cells doesn’t ensure the same
behavior at every experiment.


Furthermore, stem cells allow to develop

more phenotypes (not only neurons but also astrocytes
and oligodendrocytes) that ensure the correct contribution of

trophic substances and cellular
junctions for a better and more physiological functionality of neurons in culture.


Their multipower and ex
treme plasticity leads us to believe that interesting results could be drawn
by their organized stimulation. Stem cells grow into adult neurons in about 1 month, developing
all the essential properties of functional CNS neurons [59].

On the other side, the

culture of human neural stem cells is very delicate and

different both from
the culture of animal stem cells and of neural brain slices, requesting an extremely specialized
know
-
how.


Therefore our culture method on MEA is quite different from those repo
rted in

the previous
paragraph. Nevertheless, the culture method adopted in our experiments has been well established
in time by Prof. Angelo Vescovi’s team [25].

Cells are plated at a density of 3500 cells/cm2 in suspension in a chemically defined, serum
free
medium containing 20 ng/mL of human recombinant epidermal growth factor (EGF) and 10
ng/mL of fibroblast growth factor (FGF
-
2). After 3
-
5 days

the cultures are harvested and the cells
are mechanically dissociated and replated under the same condition
s. Our experiments have been
performed four weeks after seeding our NSCs onto MEA surfaces previously coated with adhesive
substrates, like mouse Laminin (2 ng/ml)and human Fibronectin (2 ng/ml).

Each experiment requires the culture of several MEAs , and
the correct growth and

adherence to

6

the electrodes are daily controlled. Measures are

performed only on the MEAs which present
optimal adhesion and

correct growth.




In order to ensure that the collected signals are

actually

due to electrophysiological

functionalities of neurons, every experiment is equipped

by measures from control basins with
cultured fibroblasts to compare behavior discrepancies. Moreover, at the

end of the experiments
the cultures are injected with Tetrodotoxin (TTX) , neurotoxin ab
le to abolish action potentials.
Then TTX is rinsed away (3 rinses x 5') and after 5 minutes the same measures are repeated . All
the control procedures confirmed the presence of neural electrical activity.



2.2 The hardware


The problem of the junction b
etween neuron and electrode is crucial: materials must be
biocompatible with the culture environment , and neurons must firmly adhere to the electrodes in
order to get maximum conductibility.

Our supports are constituted by glass dishes with 96 tungsten m
icroelectrodes (Fig. 1) .

Each electrode is connected, by means of a sharp insulated track, to a pad suitable for the external
connection.


Fig. 1


Portion of the MEA support


7

The distance between electrodes varies between 100 and 200

m, whereas the
diameter of each
electrode is around 20

m. During the experiments only distant electrodes have been used, thus
the actual distance between electrodes ranged between 300 and 500

m.


On the MEA four basins suitable for cell culture are installed, in such a

way as to realize more
experiments simultaneously.


From the 96 electrodes some have been chosen as neural input/output, others as ground.

The electrodes have been connected to realize special layouts, as described in the next paragraph.

The block dia
gram of the hardware is represented in Fig. 2.


Fig. 2


Block diagram of the hardware : Acquisition and stimulation systems


The whole circuit is composed by:

-

Electrodes shared among four basins in which cells are cultured

-

A shielded cable used to
send and receive electrical signals between electrodes and stimulation
circuit

-
Two shielded cables used to send analog signals to the DAQ and to transfer the digital signals
coming from the I/O ports of the DAQ to the stimulation circuit.


8

-
An electronic s
timulation circuit whose task is to convert digital 8
-
bit signals (patterns),
generated by the software resident on the PC, from electrical signals generated by the DAQ with a
logical level 0
-
5V, into electrical pulses with voltage and current suitable to
stimulate neurons.

The stimulation occurs with a 35 mV positive voltage. In order to depolarise the culture liquid,
before every bit a negative

35 mV pulse is emitted. The pulse length is 10% of the whole bit
duration, thus the whole pulse is composed by

10% negative voltage, 90% positive voltage.

As we stimulated the cells with 40 Hz frequency, the whole pulse duration is 25 ms.


The computer’s task is to generate the patterns and store on the hard disk the data coming from the
cells.

The sampling rate
is 80 Hz with an input sensibility of 61 mV. Besides, a 50 Hz notch filter has
been inserted ( European electrical supply frequency).


To avoid interference , the stimulation circuit disconnects the voltage generator when the
acquisition card gets ready to

receive the signals coming from the culture basins.

The electronic circuit is included in a plastic box whose walls have been treated with special
varnishes that efficiently shield possible EMI noise.

All the cables used for the connection between cult
ure basins, stimulation circuit and acquisition
card have been carefully shielded. We also minimized the power supply ripple using a condenser
with low ESR, in order to avoid a possible ripple in the generated signals.


The DAQ is an external device,
connected to the PC through the USB port, that can be located up
to 5 meters from the PC without batteries. This configuration allows to shorten the connections
neurons/acquisition card and to reduce at the same time the risk of possible electrical noise
generated by the computer electronic circuits (Fig. 3).


9


Fig. 3
-

Experimental setup: Microscope with videorecorder , cells in incubator and DAQ



The features of the acquisition card, produced by IOtech , Inc., include h
igh
-
resolution, 22
-
bit A/D
conver
ter, and digital calibration, frequency measurements up to 1 MHz and optical isolation
from PC. It allows programmable inputs from

31 mV to

20V full scale and it is
expandable up to 80 channels of analog and digital I/O [30].



2.3 The artificial architecture


The microelectrodes are connected following the architecture of classical Artificial Neural
Networks [40, 53, 54, 27 and
28].

On the MEAs we implemented two kinds of artificial architectures: a Kohonen Self
-
Organizing
Map [35] and a Hopfield network [29].


We chose these models due to their straightforward architecture and their resemblance to some
neurophysiological struct
ures.



10

2.3.1 The Self Organizing Map


The Kohonen Network (Self Organizing Map, SOM) has been developed in the eighties by T.
Kohonen on the basis of previous neurophysiological studies [55]: it has been shown that, on the
human cortex , cortical maps for
m themselves by self
-
organization , in such a way that near
neurons are activated by similar perceptive stimuli.

The SOM is a non
-
supervised network, i.e. it works without need of presentation of known
examples.

The network structure consists of an inp
ut layer and a so
-
called competitive layer with N
neurons. Each of them receives
n

signals
x
1
,…,x
n
coming from the n elements of the input layer,
following connections with
w
ij

weight (Fig. 4).


Fig. 4


Kohonen Network (SOM)


The intensity
I
of the

element
i

is calculated by



I
i
= D(w
i
, x)


where
D(w
i
, x)

is some distance function, for example the Euclidean one, between the input and
each neuron of the competitive layer.


11

The learning phase (Winner Take All law, WTA) consists of a competition to

evaluate which
element has the minimum input intensity, i.e. which
w
i

is the nearest to
x

.

The weights are modified following the law


w
inew
= w
iold
+
η

(x
-
w
iold
)z
i



where

0<
η
<1
is the learning rate
.


In this way the network moves more and more towards the nearest stimuli, ideally up to overlap
them: the SOM performs a mapping from a multidimensional space to a space with less
dimensions, preserving t
he starting topology: in other words, it classified a pattern as the nearest
among a set of reference elements.


2.3.2 The Hopfield Network


From 1982 to 1985 the physicist J.J. Hopfield presented a neural network model with an
interesting architecture f
rom the neurodynamical point of view.

The Hopfield network allows to memorize vectors and to recall them later.

The network consists of n fully connected neurons, as shown in Fig. 5.


Fig. 5
-

Hopfield Network



12


The inputs are applied simultaneously to
all the N nodes. Thus the weights are symmetrical and
follow the law



w
ij
=





for


i


j



= 0





for


i=j, 0


i , j


N
-
1


In the learning phase each output behaves as input on the same neuron: the new value is
establi
shed by the function


f(x
i
)= x
i

if

= T
i
(
T
i



0
is a threshold)


f(x
i
) = +1 if
> T
i




f(x
i
) =
-
1 if
< T
i



We can see an input pattern as a point in the stat
e space, that,while the network is running , moves
towards the minima representing the steady states of the network, in the lowest points of its
attraction basins [60] .When the network stops, the weights values are the network output.


If we associate to
the network an energy function E


E(x) =
-
1/2
w
ij

x
i

x
j


13


this value decreases monotonically in time .

In fact, being in this case


E/

x
i

=
-


w
ij

x
j



if


x
i

>0
w
ij

x
j

>0

if


x
i

<0
w
ij
x
j

<0




i.e. always

E


0 .


Thus after a number of iterations the network stabilizes into a minimum energy
state.

Each minimum corresponds to a stored pattern.

The m memorized forms
s
1
,...,s
m

correspond to the local minima of the E(s) function.

Therefore if we present to the network a vector s


slightly different from the stored patterns s , the
network dynam
ics will relax on the local minimum nearest to
s


, i.e.
s

.



2.4 The hybrid networks


The Artificial models have been implemented on the MEAs, culturing the stem cells on the
connection sites.
The number of cells adhering on the electrodes are 1
-
3 for
each electrode.

1)

The Kohonen network is implemented by 8 electrodes that constitute the input
layer, and three electrodes that constitute the competitive layer. The output signals are

14

collected directly from the competitive layer. The input and the competi
tive electrodes are
connected following the classical Kohonen architecture.

2)

The Hopfield network is set up by 8 completely interconnected electrodes,
following the classical Hopfield architecture. The electrodes act both as input and as
output (Fig. 6).



Fig. 6


Connections for the Kohonen Network (left) and the Hopfield Network (right) on MEA


These architectures have been simulated by software artificial neural networks and they have
shown to constitute the minimum configuration suitable to correctly
recognize two different input
bitmaps ( patterns ), namely a “0” and a “1” constituted by 9
-
element bitmaps . The central value
of the bitmap has been always considered null for sake of electronic simplicity, obtaining 8
-
element bitmaps.

The patterns have

been delivered to the hybrid networks as a train of electrical pulses in such a
way as to represent every black square of the bitmap (see Fig. 7) as a 35 mV pulse (similar to the
natural action potential) and every white square as a 0 mV pulse.


15


Fig. 7



Wave form of the stimulation pulse


Artificial neural networks and the human brain are able to recognize not only sharp images, but
also images affected by noise: for this purpose we delivered to the hybrid networks also noise
-
affected patterns (Fig. 8

and Fig. 9) in order to verify their ability to recognize them correctly.


Fig. 8


“0” bitmap and “0 with noise” bitmaps submitted to the hybrid network


Fig. 9


“1” bitmap and “1 with noise” bitmaps submitted to the hybrid network


16

Stimulation is per
formed with 40 Hz frequency. The DAQ sampling is 80 Hz for each channel. A
40 Hz low pass filter has been applied off
-
line before the signal analysis in order to highlight
possible correlations between signals at that frequency. In fact in the last decade

several studies
[42, 16 and 44] have supported the hypothesis that electrical correlated activity at 40 Hz (gamma
oscillations) frequency flows along the cortical neurons with the possible aim (or effect) to create
a functional binding of perceptions:
thus 40 Hz is the candidate frequency for evolved
functionalities in brain.

In order to allow the choice of different “0” and “1” patterns, pure or affected by noise, we
developed a software that interfaces with the acquisition card and allow to draw the

desired
bitmaps to be sent to the network . The software also allows to set the number of cycles of the
chosen sequence of patterns, the waiting time between sequences, the number of iterations, the
name of the file that will be recorded .

The output si
gnals have been analysed to evaluate the possibility of organized reaction by the
natural neurons.

Although stem cells show the essential properties of functional CNS neurons, their detailed
behavior is still in course of study [38, 41 and 26]. For this
reason our analysis does not search for
known features and artefacts in signals, but utilizes tools able to measure the degree of
organization of signals during and after training.

To this purpose after the experiment the output signals have been analysed

using the Recurrent
Quantification Analysis

[64,65] .







17

2.5 Recurrence Quantification Analysis


Recurrence Quantification Analysis (RQA) is a new quantitative tool that can be applied to time
series reconstructed with delay
-
time embedding. RQA is in
dependent of data set size, data
stationarity, and assumptions on statistical distributions of data.

RQA gives a local view of the series behaviour, because it analyses distances of pairs of points,
not a distribution of distances. Therefore, unlike auto
correlation, RQA is able to analyse fast
transients and to localize in time the features of a dynamical variation: for this reasons RQA is
ideally suited for physiological systems.


The Recurrent Plots show how the vectors in the reconstructed space are n
ear or distant each other.

The observation of recurrent points consecutive in time (forming lines parallel to the main
diagonal) is an important signature of deterministic structure. In fact the length of the longest
diagonal (recurrent) line accurately co
rrespond to the value of the maximum Lyapounov exponent
of the series. The Lyapounov exponent is a measure of chaoticity , quantifying the mean rate of
divergence of neighbouring trajectories along various directions in the phase space of an embedded
time
series. Time series of chaotic systems have a positive maximum Lyapounov exponent.


The Recurrent Plots version we adopted [36] calculates the Euclidean distances between all the
vector pairs and translates them into colour bands. Hot colours (yellow, red
, orange) are
associated to short distances between vectors, cold colours (blue, black) show long distances.
Signals repeating fixed distances between vectors are organized, signals without repeating
distances are not. In this way we obtain uniform colour
distribution for random signals, but the
more deterministic and self
-
similar is the signal, the more structured is the plot.





18

2.6 ITSOM analysis


Another kind of analysis has been performed using a novel Artificial Neural Network called
ITSOM (Inductive

Tracing Self Organizing Map) useful to highlight structures in the temporal
series of a signal.

As we have seen in paragraph 2.3.1, the main SOM feature is to identify a winning neuron
which should classify the input stream. But
two main reasons exist tha
t limit the SOM's
performances in case of strictly non
-
linear and time
-
variant input.

The first reason is that if the input topology is too tangled, the competitive layer is not able to
unfold itself enough to simulate the input topology.

The second reaso
n concerns the SOM's convergence conditions, that exist but are not easily
verifiable. Due to the nature of the SOM's output (non homologous to the input), it is not
possible to settle either a network error for each epoch, or the number of epochs after t
hat the
network training has to be stopped.

Nevertheless, as in many cases even after several thousands of epochs the convergence was not
reached, the processing time was verified to become too long for a real
-
time application.


Another problem of the SO
M, typical of any clustering algorithm, is the lack of output
explication. Once obtained a classification, the user must analyze it, comparing it to the input
values in order to extrapolate a significant output.


Thus we proposed the structural modificat
ion of the SOM called Inductive Tracing Self
-
Organizing Map (ITSOM). The dynamical properties of artificial neural networks and of the
SOM in particular are well known [52, 43, 31, 50, 51 and 14].

During simulations carried on with the SOM algorithm we ha
ve observed that, even if the
winning weights may vary at any presentation epoch, their temporal sequence tends to repeat
itself [47].


19

A deeper analysis has shown that such a sequence, provided to keep the learning rates steady
(instead of gradually decre
asing them), constitutes chaotic attractors that repeat themselves
“nearly” exactly in time with the epochs succeeding, and that, once codified by the network,
univocally characterize the input element that has determined them.

Actually the learning rule m
akes it possible for winning weight to represent an approximation
of the input value. At every epoch the new winning weight, together with the previous winner,
constitutes a second order approximation of the input value. At the n
-
th epoch, the set of n
win
ning weights represents an n
-
order approximation of the input value.

In this way, due to the countless variety of possible combinations among winning neurons, the
configurations allow to finely determine the correct value, even in the case of tangled inpu
t
topologies, despite of the small number of competitive neurons and their linear topology.

In the following step the network performs a real induction process, because after a many
-
to
-
few vector quantization from the input to the weight layer (to be prec
ise, to the chaotic
configurations of winning weights), a few
-
to
-
many procedure is performed from the chaotic
configurations corresponding to the input set (Fig. 10) codified by the network.


Fig. 10


The ITSOM Network identifies a series of winning neu
rons in time



20

It should be stressed that the ITSOM's crucial feature is that the network does not need to be
brought to convergence, as the cyclic configurations stabilize their structure within a small
number of epochs, then keep it steady through time.


After interrupting the network processing phase, an algorithm is needed that codifies the
obtained chaotic configurations into a small set of outputs.

The algorithm which has shown best performances and computational load among the tested
pattern recognit
ion algorithms is based on a z
-
score calculus .

The cumulative scores related to each input have been normalized following the distribution of
the standardized variable z given by




z = (x
-



)/



where


is the average of the scores on all the competitive layer weights and


is the mean
squared deviation. Once fixed a threshold 0<

<=1 , we have put





z = 1 for z>




z = 0 for z




In this way every winning configuration is repr
esented by a binary number with as many 1 and
0 as many the competitive layer weights.

Then the task of comparing these binary numbers is straightforward.

It has been verified that the


threshold size is not critical: fixing it to 0.5 we have obtained t
he
best results with any input stream.


21

The z
-
score method has shown to be steady with regard of the performances, and
computationally not expensive, being linear in the number of the competitive layer weights.

But it is worth emphasizing that the z
-
score
algorithm allows the network to reach its best
performances in a very small number of epochs (often less then 15).

This allows the network to complete its work within an insensible time, and to actually assert
the possibility of a real
-
time processing.

The

good performances of this feedback network have been tested in the classification of
neurological diseases [45], for equalization and demodulation of GSM signals [15] , for image
classification [46] and for EEG analysis [44] .

The small computational lo
ad is an element that makes the ITSOM suitable for a possible
hardware implementation [10, 11].



3. RESULTS


During the experiments cells are kept in a controlled environment a 37 °C. At the end of the
experiment the neurons are alive and maintain their

functionalities.

In order to check if the signals received by the acquisition device were actually coming from
neurons, we measured the reactions of the only culture liquid with fibroblasts (Fig. 11) ,


Fig. 11
-

Culture liquid output during stimulatio
n with “0” patterns: conductor
-
like behavior


22


comparing them with those coming from the cells (Fig. 12).


Fig. 12


Network output during stimulation with “0” patterns: low
-
voltage behavior


It is evident that the network reacts to the “0” pattern, consti
tuted by the highest voltage
(11111111), emitting the lowest voltages. The culture liquid, instead, answers to the “0” pattern
with a high voltage, as expected by a conductive medium .

In Fig. 13 is depicted the reaction of the Kohonen network after stimul
ation with “0” patterns,
pure and affected by noise (green circles) , and with “1” patterns (red circles), pure and affected
by noise.


Fig. 13


Kohonen Network output during training: different voltages for different patterns



23

Similar effects have been
shown by the Hopfield network. At the end of the experiment we
measured the network output in order to evaluate if the neurons had “stored” information in some
way. Differently from the only culture liquid, that shows the same behaviour before and after
ex
periment, the network output retains different voltages .


The analysis of our data using the RQA method lead to interesting results. Signals coming from
similar bitmaps gave rise to similar Recurrent Plots.

Moreover, in the following figures you can see t
he self
-
organization of a single output channel
before stimulation, during the training, during the testing phase and after stimulation.

Fig. 14a shows one output channel of the Kohonen network before stimulation. Colours are cold
and unstructured, showing

lack of self
-
organization.

The training phase shows a change in the structure. The Recurrent Plot of the channel during the
testing phase shows wide uniform hot colour bands corresponding to a high organization. Fig. 14b
is the plot of the output channel
after the end of stimulations. In this case the uniform bands
further widen, showing that the signal remains self
-
organized in time.


a






b

Fig. 14
-

RQA plot of the Kohonen network before stimulation (a, no organization) and after training(b, high or
ganization)



24

This analysis shows that introduction of organized stimuli modifies the network structure and
increases the information content even after the end of stimulation, suggesting a form of learning
and memorization..

We applied the same procedure t
o the output signals coming from the Hopfield network.

Fig. 15a shows one channel after stimulation with the “0” pattern :we see wide organized bands
with peculiar features, different from the other channels. Fig. 15b shows the same channels after
stimulat
ion with “1” pattern.





a






b

Fig. 15
-

RQA plot of one channel of the Hopfield network after stimulation with “0” patterns (a) and “1” patterns (b)


This analysis shows that the network behaves differently depending on the input signal and on the
d
ifferent channels.


In order to confirm the results drawn by means of the RQA analysis , and to specify the
information content of the output signals, we used the ITSOM artificial neural network described
in 2.6 .

The ITSOM network has been applied to sig
nals coming from outputs of different bitmaps.

We used a network with 250 input neurons and 15 competitive neurons. Due to organized content
of the signals, the chaotic attractors repeated themselves after around 20 epochs.


25


As expected after the RQA analy
sis, the networks has discriminated different bitmaps with
different series of winning neurons , whereas similar bitmaps have shown an identical series of
winning neurons

(see Tab. 1).










Tab. 1


z
-
score code of some “0” patterns : the ITSOM netwo
rk assigns similar patterns the same code



4.

DISCUSSION AND CONCLUSIONS


After our analysis of the output signals of the networks we can reasonably affirm that the networks
show an organized behavior after the stimulation with patterns, and they are able to

answer
selectively to different patterns. The signal behavior changes depending on the network channels,
and similar patterns give rise to similar answers.

Thus we can say that the networks have shown a form of selective coding, highlighting a strong
self
-
organization as a reply to stimulation,
that persists for a long time after the stimulation bursts.


Moreover, despite of the lack of a detailed neurophysiological interpretation of the signal tracing,
the ITSOM network has allowed to distinguish the d
ifferent information contents of the signals.
We could show that similar patterns give rise to output signals containing similar chaotic attractors

0


pattern


type 1


1


0


0


0


1


1


0


0


1


1


0


0


1


1


0


0


1



0


pattern


type 2



1


0


0


0


1


1


0


0


1


1


0


0


1


1


0


0


1



0


pattern


type 3



1


0


0


0


1


1


0


0


1


1


0


0


1


1


0


0


1



0


pattern


type 4


1


0


0


0


1


1


0


0


1


1


0


0


1


1


0


0


1



0


pattern


type 5




1


0


0


0


1


1


0


0


1


1


0


0


1


1


0


0


1



0


pattern


type 6




1


0


0


0


1


1


0


0


1


1


0


0


1


1


0


0


1



0


pattern


type 7



1


0


0


0


1


1


0


0


1


1


0


0


1


1


0


0


1



0


pattern


type 8



1


0


0


0


1


1


0


0


1


1


0


0


1


1


0


0


1



0


pattern


type 9




1


0


0


0


1


1


0


0


1


1


0


0


1


1


0


0


1



26

which have been codified, whereas different pattern lead to attractors corresponding to different
codes.


Being able to discriminate ed interpret the information content of the biological network, we are
planning to use in the future these outputs in several ways.

In fact, aim of this kind of research is on one side to improve the knowledge of the
neurophysio
logical learning and memory functionalities; on the other side it would be possible to
evaluate the feasibility of a hybrid electronic
-
biological device, conceiving the possibility of
biological computation, or of non
-
invasive neurological prostheses, able

to improve or substitute
damaged nervous functionalities .


The culture method we adopted revealed to be suitable for our experiments, ensuring the necessary
survival of cells, but at the moment neurons don’t keep alive more than around two months.
Nonet
heless better culture method and more suitable MEA supports are under study and a future
improvement of the culture duration is expected.

Another problem that should be solved is the increase of complexity of the artificial connection on
the MEAs, in ord
er to pass from prototype patterns to complex patterns.

By increasing the number of connections and of input
electrodes, the acquisition card will share
the sampling rate into more channels , diminishing its performances. At the moment we have
already acqu
ired a new more powerful DAQ (National Instruments
16 bit

PCI
-
6036E), and we
started experiments with six more complex input bitmaps, maintaining the same number of input
channels and delivering the patterns to the hybrid network 8 bits at a time, followi
ng a well
-
known Artificial Neural Networks technique .

Analyses of the results are under way, but preliminary studies show the same recognition
capabilities as those described in this paper.



27

Another improvement will be the implementation of a printed ci
rcuit board instead of the current
wired hardware, in order to ensure robustness and flexibility .


As we verified the possibility to codify the neuron output, we are developing the interface
between the codified output and a simple actuator (a minirobot),

with the purpose to experiment a
complete chain perception (input patterns)


hybrid neural network


action : this is possible due
to the ITSOM real time coding of the output stream.

Off
-
line experiments have been already carried out and the whole real
-
time system is expected in
the next few month.



ACKNOWLEDGEMENTS


We are strongly indebted to Prof. G. Degli Antoni (University of Milan) for his valuable
suggestions and encouragement , to Dr. Francesca Gregori for her substantial contribution and to
ST

Microelectronics for the important support.




REFERENCES


1.

Akin T., Najafi K., Smoke R.H. and Bradley R.M., A micromachined silicon electrode for
nerve regeneration applications,
IEEE Trans Biomed Eng 41

(1994) 305
-
313 .

2.

Bels B. and Fromherz P., Transis
tor array with an organotypic brain slice: field potential
records and synaptic currents,
European Journal of Neuroscience 15
, (2002) 999
-
1005.


28

3.

Bonifazi P. and Fromherz P., Silicon Chip for Electronic communication between nerve
cells by non
-
invasive int
erfacing and analog
-
digital processing,
Advanced Material 17

(2002).

4.

Borkholder D.A., Bao J, Maluf N.I., Perl E.R. and Kovacs G.T., Microelectrode arrays for
stimulation of neural slice preparations ,
J. Neuroscience Methods 7

(1997) 61
-

66.

5.

Bove M., Mar
tinoia S., Grattarola M. and Ricci D., The neuron
-
transistor junction:
Linking equivalent electric circuit models to microscopic descriptions,
Thin Solid Films
285

(1996) 772
-
775.

6.

Braun D. and Fromherz P., Fast Voltage Transients in Capacitive Silicon
-
to
-
Cell
Stimulation Detected with a Luminescent Molecular Electronic Probe,
Physical Review
Letters 13, 86

(2001) 2905
-
2908.

7.

Breckenridge L.J., Wilson R.J.A., Connolly P., Curtis A.S.G., Dow J.A.T., Blackshaw S.E.
and Wilkinson C.D.W., Advantages of using m
icrofabricated extracellular electrodes for in
vitro neuronal recording,
J Neuroscience Research 42

(1995) 266
-

276.

8.

Canepari M, Bove M., Mueda E., Cappello M., Kawana A., Experimental analysis of
neural dynamics in cultured cortical networks and transitio
ns between different patterns of
activity,
Biological Cybernetics 77

(1997) 153
-
162.

9.

Carmena, J.M., Lebedev, M.A., Crist, R.E., O'Doherty, J.E., Santucci, D.M., Dimitrov,
D.F., Patil, P.G., Henriquez, C.S., & Nicolelis, Learning to control a brain
-
machine
interface for reaching and grasping by primates,
M.A.L. PLoS 1

(2003) 193
-
208.

10.

Chen O., Sheu B., Fang W., Image Compression on VLSI Neural
-
based Vector
Quantizer,
Information Processing and Management 28, 6

(1992).

11.

Choi J., Bang S.H., Sheu B.J., A Programm
able Analog VLSI Neural Network
Processor for Communication receivers,
IEEE Trans on Neural Network. 4, 3

(1993).


29

12.

De Marse T.B., Wagenaar D.A., Potter S.M., The neurally
-
controlled artificial animal: a
neural computer interface between cultured neural net
works and a robotic body”,
SFN
2002
, Orlando, Florida (2002).

13.

Egert U., Schlosshauer B., Fennrich S., Nisch W., Fejtl M., Knott T., M
ü
ller T. and
Hammerle H., A novel organotypic long
-
term culture of the rat hippocampus on substrate
-
integrated microelectr
ode arrays,
Brain Resource Protoc 2

(1988) 229
-
242.

14.

Ermentrout B., Complex Dynamics in WTA neural Networks with slow inhibition,
Neural Networks 5

(1992) 415
-
431.

15.

Favalli L., Mecocci A., Pizzi R., A soft receiver using recurrent networks,
Proc.
EUSIPCO96
, VII European Signal Processing Conference
, Trieste (1996).

16.

Freeman W.J., Relation of olfactory EEG on behaviour: time series analysis,
Behavioural
Neuroscience 100

(1987).

17.

Fromherz P, Electrical Interfacing of Nerve Cells and Semiconductor Chips,
Chemph
yschem 3
, (2002) 276
-
284.

18.

Fromherz P. , Muller C. O. , Weis R., Neuron
-
Transistor: electrical transfer
function
measured by the Patch
-
Clamp technique,
Phys. Rev. Lett. 71
(1993) 4079
-
4082.

19.

Fromherz P. and Schaden H., Defined neuronal arborisations by guid
ed outgrowth of leech
neurons in culture,
Eur J Neuroscience 6

(1994).

20.

Fromherz P. and Stett A.,
Silicon
-
Neuron Junction: Capacitive Stimulation of an
Individual Neuron on a Silicon Chip
( The American Physical Society ,1995).

21.

Fromherz P., Electrical Interf
acing of Nerve Cells and Semiconductor Chips,
Chemphyschem 3

(2002) 276
-
284.

22.

Fromherz P., Offenhäusser A., Vetter T., Weis J.,

A Neuron
-
Silicon
-
Junction: A Retzius
-
Cell of the Leech on an Insulated
-
Gate Field
-
Effect Transistor
,

Science 252

(1991)

1290
-
129
3
.


30

23.

Garcia P.S., Calabrese R.L., DeWeerth S.P., Ditto W. , Simple Arithmetic with Firing Rate
Encoding in Leech Neurons: Simulation and Experiment,
Proceedings of the XXVI
Australasian computer science conference

16,

Adelaide,


(2003) 55


60.

24.

Gritti A G
alli R. Vescovi A.L. , Culture of stem cell of central nervous system, Federoff
ed.,( Humana Press III 2001) 173
-
197.

25.

Gritti A., Frolichsthal
-
Schoeller P., Galli R., Parati E.A., Cova L., Pagano S.F., Bjornson
C.R., Vescovi A., Epidermal and fibroblast gr
owth factors behave as mitogenic regulators
for a single multipotent stem cell
-
like population from the subventricular region of the
adult mouse forebrain,
J. Neurosci. 19, 9

(1999) 3287
-
97.

26.

Gritti A., Rosati B., Lecchi m., Vescovi A.L. and Wanke E., Excit
able properties in
astrocytes derived from human embryonic CNS stem cells,
European Journal of
Neuroscience 12
, (2000) 3549
-
3559.

27.

Haken H.,
Synergetic computers and cognition (A top
-
down approach to neural nets)

(Springer 1991).

28.

Haykin S.,
Neural networks
(A comprehensive foundation)
( MacMillan Coll. Pub. 1994).

29.

Hopfield J.J., Neural Networks and Physical Systems with Emergent Collective
Computational Abilities,
Proc. Nat.
Acad.
Sci US, 81

(1984).

30.

IOTech , Personal Daq User’s Manual USB Data Acquisition Mo
dules ,

http://www.iotech.com/catalog/daq/persdaq.html

(2001),

31.

Jeffries C.,
Code Recognition and Set Selection with Neural Networks

(Birkhauser
Boston 1991).

32.

Jenkner M and Fromherz P, Bistabi
lity of membrane conductance in cell adhesion
observed in a neuron transistor,
Phys Rev Lett 79

(1997) 4705
-
4708.

33.

Jenkner M, Bert Muller, Peter Fromherz, Interfacing a silicon chip to pairs of snail neurons
connected by electrical synapses,
Biological Cy
bernetics 84
, (2001) 239
-

249.


31

34.

Jimbo Y. and Robinson H.P.C., Propagation of spontaneous synchronized activity in
cortical slice cultures recorded by planar electrode arrays,
Bioelectrochemistry 5

(2000)
107
-
115.

35.

Kohonen T.,
Self
-
Organisation and Associati
on Memory

(Springer Verlag 1990).

36.

Kononov E.,
http://home.netcom.com/~eugenek/download.html
, (1996).

37.

Lindner J.F,.
Ditto W., Exploring the nonlinear dynamics of a physiologically viable
model n
euron,
AIP Conf. Proc. 1

(1996) 375.

38.

Magistretti J., Mantegazza M. Guatteo E., and Wanke E., Action potential recorded with
patch clamp amplifiers: are they genuine ?,
Trends Neurosci. 19
(1996) 531
-
534.

39.

Maher M.P., Pine J., Wright J. and Tai Y.C., The ne
urochip: a new multielectrode device
for stimulating and recording from cultured neurons,
Neuroscience Methods 87

(1999) 45
-
56.

40.

Mc Culloch W. e Pitt W., A Logical Calculus of the Ideas Immanent in Nervous Activity,
Bulletin of Mathematical Biophysics 5

(1
943) 115
-
133.

41.

McKay R.D.G. , Stem cells in the central nervous system,
Science 276

(1997) 66
-
71.

42.

Menon V., Freeman W.J., Spatio
-
temporal Correlations in Human Gamma Band
Electrocorticograms,
Electroenc. and Clin.. Neurophys. 98

(1996) 89


102.

43.

Pineda F.J.
, Dynamics and Architectures for Neural Computation,
Journal of
Complexity 4

(1988) 216
-
245.

44.

Pizzi R., de Curtis M., Dickson C., Evidence of Chaotic Attractors in Cortical Fast
Oscillations Tested by an Artificial Neural Network, in: J. Kacprzyk ed.,
Advan
ces in Soft
Computing,
Physica Verlag (2002).

45.

Pizzi R., Reni G., Sicurello F., Varini G., Comparison between clustering algorithms
and neural networks for the classification of neurological diseases,
Proc. Softstat 97
,
Heidelberg (1997).


32

46.

Pizzi R., Sicure
llo F., Varini G., Development of an inductive self
-
organizing network for
the real
-
time segmentation of diagnostic images ,
Proc. 3th International Conference of
Neural Networks and Expert Systems in Medicine and Healthcare

Pisa

(1998) 44
-
50.

47.

Pizzi R., Th
eory of Dynamical Neural Systems with application to
Telecommunications, PhD Dissertation, University of Pavia (1997).

48.

Potter S.M., Distributed processing in cultured neuronal networks, in:
Progress in
Brain
Research
,
M.A.L. Nicolelis ed.
,
130

(
Elsevier S
cience B.V. 2001) .

49.

Reger, B, Fleming, KM, Sanguineti, V, Simon Alford, S, Mussa
-

Ivaldi, FA. ,Connecting
Brains to Robots: An Artificial Body for Studying the Computational Properties of Neural
Tissues,
Artificial Life 6

(2000) 307
-
324.

50.

Ritter H., Schulte
n K., Convergence properties of Kohonen’s Topology Conserving
Maps

: Fluctuations, Stability, and Dimension Selection,
Biological Cybernetics 60
,
(
1988) 59
-
71.

51.

Ritter H., Schulten K., On the Stationary State of Kohonen’s Self
-
Organizing Sensory
Mapping,
Bi
ological Cybernetics 54

(1986) 99
-
106.

52.

Rosen R., Dynamical System Theory in Biology,
Stability Theory and its applications 1
,
(Wiley Interscience 1970).

53.

Rosenblatt F., The perceptron: a probabilistic model for information storage and
organization in the
brain,
Psychological Review 65

(1958) 386
-
408.

54.

Rumelhart, D. E., Mcclelland, J. L. ,
Parallel Distributed Processing
, (MIT Press,
Cambridge 1986).

55.

Saarinen J., Kohonen T., Self
-
organized formation of colour maps in a model cortex,

Perception, Feb

( 1985)
.

56.

Schatzthauer R.a nd Fromherz P., Neuron
-
silicon junction with voltage gated ionic
currents,
European Journal of Neuroscience

10

(1998) 1956
-
1962.


33

57.

Schiff S.J.,Jerger K., Duong D.H., Chang T., Spano M.L. and Ditto W.,

Controlling chaos in the brain,
Natu
re 8, 25

(1994).

58.

School M., Sprössler C., Denyer M., Krause M., Nakajima K., Maelicke A., Knoll W.,
Offenhäusser, Ordered networks of rat hippocampal neurons attached to silicon oxide
surfaces,
Neuroscience Methods 104

(2000) 65
-
75.

59.

Song H., Stevens C.F.

and Gage F.H.,
Neural stem cells from adult hippocampus develop
essential properties of functional CNS neurons,
N
ature Neuroscience 5, 5

(2002)
.

60.

Tank D.W.
and Hopfield J.J. ,
Neural architecture and biophysics for sequence
recognition in Neural Models o
f Plasticity

(Academic Press l989).

61.

Vescovi A.L., Parati E.A. Gritti A. Poulin P., Ferrario M., Wanke E., Frölichsthal
-
Schoeller P., Cova L., Arcellana
-
Panlilio M., Colombo A., and Galli R., Isolation and
cloning of multipotential stem cells from the embry
onic human CNS and establishment of
transplantable human neural stem cell lines by epigenetic stimulation. ,
Exp. Neurol. 156
,
(1999) 71
-
83.

62.

Weis R., Müller B., Fromherz P.,
Neuron Adhesion on Silicon Chip probed by an Array of
Field
-
Effect Transistors
,

Ph
ys.Rev.Lett. 76

(1996) 327
-
330.

63.

Wilson, R.J., Breckenridge L., Blackshaw S.E., Connolly P., DowJ.A.T., Curtis A.S.G.,
and Wilkinson C.D.W., Simultaneous multisite recordings and stimulation of single
isolated leech neurons using planar extracellular electr
ode arrays,
Neuroscience Methods
53

(1994) 101
-
110.

64.

Zbilut J.P., Webber C.L., Embeddings and delays as derived from quantification of
recurrent plots

Phys. Lett. 171

(1992).


65.

Zbilut J.P., Thomasson N., Webber Jr C.L. , Recurrence quantification analysis as

a tool
for nonlinear exploration of nonstationary cardiac signals,
Medical Engineering and
Physics 24

(2002) 53
-
60.


34

66.

Zeck G. and Fromherz P., Noninvasive neuroelectronic interfacing with synaptically
connected snail neurons on a semiconductor chip,
Proceed
ings of the National Academy of
Sciences 98

(2001) 10457
-
10462.