<html>
  <head>
    <meta content="text/html; charset=ISO-8859-1"
      http-equiv="Content-Type">
  </head>
  <body text="#000000" bgcolor="#FFFFFF">
    <div class="moz-cite-prefix">Dear Andras,<br>
      <br>
        I see why it may seem on the first look that practopoiesis is
      somehow related to reinforcement learning. However, any closer
      look to the theory will reveal that this is not the case.
      Practopoiesis is related to just a little bit to reinforcement
      learning, and it is as much related to any other brain theory or
      algorithm. At the end, in fact, it is quite a novel and exciting
      way of looking at brain function.<br>
      <br>
       Maybe the contribution of practopoiesis can be appreciated best
      if described in the language of learning. Imagine a system that
      does not have one or two general learning rules that are applied
      to all of the units of the system. Instead, imagine that the
      number of different learning rules equals the number of units in
      the system; a billion neurons means billion different learning
      rules, whereby each learning rule maximizes a different function.
      <br>
      <br>
      Moreover, imagine that, in this system, all its long-term memories
      (explicit and implicit) are stored in those learning rules. Thus,
      long-term memories are not stored primarily in the weights of
      neuron connections (as widely presumed) but in the rules by which
      the system changes its network in response to sensory inputs.
      Then, when we retrieve a piece of information, or recognize a
      stimulus, or make a decision, we use these learning mechanisms by
      quickly applying them to the network (e.g., every second or even
      faster) as a function of the incoming sensory inputs (or the
      sensory-motor contingencies). As a result, the network 
      continuously changes its architecture with very high rate, and can
      quickly come back to its previous architecture if it is presented
      with a sequences of stimuli that has been presented already
      previously. One of the key point is that this process of changing
      the network is how we think, perceive, recall, categorize, etc.<br>
      <br>
      This process requires one more set of learning mechanisms that lay
      behind those mentioned rules containing our long-term memory. This
      latter set is responsible for acquiring our long-term memories
      i.e., for determining for each unit which learning rules it needs
      to use. Thus, there is a process of learning how to learn.<br>
      <br>
      Practopoietic theory explain how this is possible, why it works,
      how such systems can be described in generalized cybernetics
      terms, and why this approach is sufficiently adaptive to produce
      intelligence on par with that of humans. In practopoietic theory,
      the learning rules that store our long-term memories are referred
      to, not as "learning", but as "reconstruction of knowledge" or in
      Greek "anapoiesis". The paper also reviews behavioral evidence
      indicating that our cognition in fact is anapoietic by its nature.<br>
      <br>
      I hope that this helps understand that practopoiesis is something
      totally new and cannot simply be described with the existing
      machine-leaning approaches and brain theories.<br>
      <br>
      With best regards,<br>
      <br>
      Danko<br>
      <br>
      <br>
      On 4/14/14 3:05 PM, Andras Lorincz wrote:<br>
    </div>
    <blockquote
cite="mid:c90c18121a2a437a9a5201342cba0c5a@DB4PR06MB173.eurprd06.prod.outlook.com"
      type="cite">
      <meta http-equiv="Content-Type" content="text/html;
        charset=ISO-8859-1">
      <style type="text/css" style="display:none"><!--P{margin-top:0;margin-bottom:0;} .ms-cui-menu {background-color:#ffffff;border:1px rgb(171, 171, 171) solid;font-family:'Segoe UI WPC', 'Segoe UI', Tahoma, 'Microsoft Sans Serif', Verdana, sans-serif;font-size:11pt;color:rgb(51, 51, 51);} .ms-cui-menusection-title {display:none;} .ms-cui-ctl {vertical-align:text-top;text-decoration:none;color:rgb(51, 51, 51);} .ms-cui-ctl-on {background-color:rgb(223, 237, 250);opacity: 0.8;} .ms-cui-img-cont-float {display:inline-block;margin-top:2px} .ms-cui-smenu-inner {padding-top:0px;} .ms-owa-paste-option-icon {margin: 2px 4px 0px 4px;vertical-align:sub;padding-bottom: 2px;display:inline-block;} .ms-rtePasteFlyout-option:hover {background-color:rgb(223, 237, 250) !important;opacity:1 !important;} .ms-rtePasteFlyout-option {padding:8px 4px 8px 4px;outline:none;} .ms-cui-menusection {float:left; width:85px;height:24px;overflow:hidden}.wf {speak:none; font-weight:normal; font-variant:n!
 ormal; tex
t-transform:none; -webkit-font-smoothing:antialiased; vertical-align:middle; display:inline-block;}.wf-family-owa {font-family:'o365Icons'}@font-face {  font-family:'o365IconsIE8';  src:url('https://r4.res.outlook.com/owa/prem/15.0.918.10/resources/styles/office365icons.ie8.eot?#iefix') format('embedded-opentype'),         url('https://r4.res.outlook.com/owa/prem/15.0.918.10/resources/styles/office365icons.ie8.woff') format('woff'),         url('https://r4.res.outlook.com/owa/prem/15.0.918.10/resources/styles/office365icons.ie8.ttf') format('truetype');  font-weight:normal;  font-style:normal;}@font-face {  font-family:'o365IconsMouse';  src:url('https://r4.res.outlook.com/owa/prem/15.0.918.10/resources/styles/office365icons.mouse.eot?#iefix') format('embedded-opentype'),         url('https://r4.res.outlook.com/owa/prem/15.0.918.10/resources/styles/office365icons.mouse.woff') format('woff'),         url('https://r4.res.outlook.com/owa/prem/15.0.918.10/resources/styles/office!
 365icons.m
ouse.ttf') format('truetype');  font-weight:normal;  font-style:normal;}.wf-family-owa {font-family:'o365IconsMouse'}.ie8 .wf-family-owa {font-family:'o365IconsIE8'}.ie8 .wf-owa-play-large:before {content:'\e254';}.notIE8 .wf-owa-play-large:before {content:'\e054';}.ie8 .wf-owa-play-large {color:#FFFFFF/*$WFWhiteColor*/;}.notIE8 .wf-owa-play-large {border-color:#FFFFFF/*$WFWhiteColor*/; width:1.4em; height:1.4em; border-width:.1em; border-style:solid; border-radius:.8em; text-align:center; box-sizing:border-box; -moz-box-sizing:border-box; padding:0.1em; color:#FFFFFF/*$WFWhiteColor*/;}.ie8 .wf-size-play-large {width:40px; height:40px; font-size:30px}.notIE8 .wf-size-play-large {width:40px; height:40px; font-size:30px}--></style>
      <div id="OWAFontStyleDivID"
style="font-size:12pt;color:#000000;background-color:#FFFFFF;font-family:Calibri,Arial,Helvetica,sans-serif;">
        <p>You are heading reinforcement learning and the
          so-called temporal difference learning in it. This is a good
          direction, since many little details can be mapped to the
          corico-basal ganglia-thalamocortical loops. Nonetheless, this
          can't explain everything. The separation of outliers from the
          generalizable subspace(s) is relevant, since the latter
          enables one to fill in missing information, whereas the former
          does not. This was the take-home message of the Netflix
          competition and the subsequent developments on exact matrix
          completion.<br>
        </p>
        <p>Andras<br>
        </p>
        <div>
          <p><br>
          </p>
          <div name="divtagdefaultwrapper"
            style="font-family:Calibri,Arial,Helvetica,sans-serif;
            font-size:; margin:0">
            <div>
              <div>
                <div>
                  <div><font>
                      <p class="MsoNormal" style="font-size:13px;
                        font-family:Calibri,Arial,Helvetica,sans-serif;
                        margin:0cm 0cm 0pt">
                        _________________________<br>
                        <font size="2">Andras Lorincz<br>
                          ECCAI Fellow</font></p>
                      <p class="MsoNormal" style="font-size:13px;
                        font-family:Calibri,Arial,Helvetica,sans-serif;
                        margin:0cm 0cm 0pt">
                        <font size="2">email: <a class="moz-txt-link-abbreviated" href="mailto:lorincz@inf.elte.hu">lorincz@inf.elte.hu</a> <br>
                          home: </font><a moz-do-not-send="true"
                          href="http://people.inf.elte.hu/lorincz"><font
                            size="2">http://people.inf.elte.hu/lorincz</font></a><font
                          size="2">  </font></p>
                    </font></div>
                </div>
              </div>
            </div>
          </div>
        </div>
        <div style="color: rgb(40, 40, 40);">
          <hr tabindex="-1" style="display:inline-block; width:98%">
          <div id="divRplyFwdMsg" dir="ltr"><font style="font-size:11pt"
              face="Calibri, sans-serif" color="#000000"><b>From:</b>
              Connectionists
              <a class="moz-txt-link-rfc2396E" href="mailto:connectionists-bounces@mailman.srv.cs.cmu.edu"><connectionists-bounces@mailman.srv.cs.cmu.edu></a> on
              behalf of Danko Nikolic
              <a class="moz-txt-link-rfc2396E" href="mailto:danko.nikolic@googlemail.com"><danko.nikolic@googlemail.com></a><br>
              <b>Sent:</b> Monday, April 14, 2014 12:51 PM<br>
              <b>To:</b> <a class="moz-txt-link-abbreviated" href="mailto:connectionists@mailman.srv.cs.cmu.edu">connectionists@mailman.srv.cs.cmu.edu</a><br>
              <b>Subject:</b> Re: Connectionists: how the brain works?
              (UNCLASSIFIED)</font>
            <div> </div>
          </div>
          <div>
            <div class="moz-cite-prefix">Dear all,<br>
              <br>
                It has been very interesting to follow the discussion on
              the functioning of ART, stability-plasticity dilemma and
              the related issues. In that context, I would like to point
              to an exciting property of the practopoietic theory, which
              enables us to understand what is needed for a general
              solution to the problems similar to the
              stability-plasticity dilemma.
              <br>
              <br>
              The issue of stability-plasticity dilemma can be described
              as a problem of deciding when a new category of a stimulus
              needs to be created and the system has to be adjusted as
              opposed to deciding to treat the stimulus as old and
              familiar and thus, not needing to adjust. Practopoietic
              theory helps us understand how  a general solution can be
              implemented for deciding whether to use old types of
              behavior or to come up with new ones. This is possible in
              a so-called "T_3 system" in which a process called
              "anapoiesis" takes place. When a system is organized in
              such a T_3 way, every stimulus, old or new, is treated in
              the same fashion, i.e., as new. The system always
              adjusts--to everything(!)--even to stimuli that have been
              seen thousands of times. There is never a simple direct
              categorization (or pattern recognition) in which a
              mathematical mapping would take place from input vectors
              to output vectors, as traditionally implemented in
              multi-layer neural networks.
              <br>
              <br>
              Rather the system readjusts itself continuously to prepare
              for interactions with the surrounding world. The only
              simple input-output mappings that take place are the
              sensory-motor loops that execute the actual behavior. The
              internal processes corresponding to perception,
              recognition, categorization etc. are implemented by the
              mechanisms of internal system adjustments (based on
              anapoiesis). These mechanisms create new sensory-motor
              loops, which are then most similar to the traditional
              mapping operations. The difference between old and new
              stimuli (i.e., familiar and unfamiliar) is detectable in
              the behavior of the system only because the system adjusts
              quicker to the older that to the newer stimuli.
              <br>
              <br>
              The claimed advantage of such a T_3 practopoietic system
              is that only such a system can become generally
              intelligent as a whole and behave adaptively and
              consciously with understanding of what is going on around;
              The system forms a general "adjustment machine" that can
              become aware of its surroundings and can be capable of
              interpreting the situation appropriately to decide on the
              next action. Thus, the perceptual dilemma of stability vs.
              plasticity is converted into a general understanding of
              the current situation and the needs of the system. If the
              current goals of the system requires treating a slightly
              novel stimulus as new, it will be treated as "new".
              However, if a slight change in the stimulus features does
              not make a difference for the current goals and the
              situation, than the stimulus will be treated as "old".<br>
              <br>
              Importantly, practopoietic theory is not formulated in
              terms of neurons (inhibition, excitation, connections,
              changes of synaptic weights, etc.). Instead, the theory is
              formulated much more elegantly--in terms of interactions
              between cybernetic control mechanisms organized into a
              specific type of hierarchy (poietic hierarchy). This
              abstract formulation is extremely helpful for two reasons.
              First, it enables one to focus on the most important
              functional aspects and thus, to understand much easier the
              underlying principles of system operations. Second, it
              tells us what is needed to create intelligent behavior
              using any type of implementation, neuronal or
              non-neuronal.<br>
              <br>
              I hope this will be motivating enough to give
              practopoiesis a read.<br>
              <br>
              With best regards,<br>
              <br>
              Danko<br>
              <br>
              <br>
              <br>
              Link:<br>
              <a moz-do-not-send="true" class="moz-txt-link-freetext"
                href="http://www.danko-nikolic.com/practopoiesis/">http://www.danko-nikolic.com/practopoiesis/</a><br>
              <br>
              <br>
              <br>
              On 4/11/14 2:42 AM, Tsvi Achler wrote:<br>
            </div>
            <blockquote type="cite">
              <div dir="ltr">I can't comment on most of this, but I am
                not sure if all models of sparsity and sparse coding
                fall into the connectionist realm either because some
                make statistical assumptions.
                <div>-Tsvi</div>
              </div>
              <div class="gmail_extra"><br>
                <br>
                <div class="gmail_quote">On Tue, Apr 8, 2014 at 9:19 PM,
                  Juyang Weng <span dir="ltr">
                    <<a moz-do-not-send="true"
                      href="mailto:weng@cse.msu.edu" target="_blank">weng@cse.msu.edu</a>></span>
                  wrote:<br>
                  <blockquote class="gmail_quote" style="margin:0 0 0
                    .8ex; border-left:1px #ccc solid; padding-left:1ex">
                    <div bgcolor="#FFFFFF">Tavi:<br>
                      <br>
                      Let me explain a little more detail:<br>
                      <br>
                      There are two large categories of biological
                      neurons, excitatory and inhibitory.   Both are
                      developed through mainly signal statistics,
                      <br>
                      not specified primarily by the genomes.   Not all
                      people agree with my this point, but please
                      tolerate my this view for now.  
                      <br>
                      I gave a more detailed discussion on this view in
                      my NAI book. <br>
                      <br>
                      The main effect of inhibitory connections is to
                      reduce the number of firing neurons (David Field
                      called it sparse coding), with the help of
                      <br>
                      excitatory connections.  This sparse coding is
                      important because those do not fire are long term
                      memory of the area at this point of time.<br>
                      My this view is different from David Field.  He
                      wrote that sparse coding is for the current
                      representations.  I think sparse coding is
                      <br>
                      necessary for long-term memory. Not all people
                      agree with my this point, but please tolerate my
                      this view for now.  
                      <br>
                      <br>
                      However, this reduction requires very fast
                      parallel neuronal updates to avoid uncontrollable
                      large-magnitude oscillations. 
                      <br>
                      Even with the fast biological parallel neuronal
                      updates, we still see slow but small-magnitude
                      oscillations such as the
                      <br>
                      well-known theta waves and alpha waves.   My view
                      is that such slow but small-magnitude oscillations
                      are side effects of
                      <br>
                      excitatory and inhibitory connections that form
                      many loops, not something really desirable for the
                      brain operation (sorry,
                      <br>
                      Paul Werbos).  Not all people agree with my this
                      point, but please tolerate my this view for now.
                      <br>
                      <br>
                      Therefore, as far as I understand, all computer
                      simulations for spiking neurons are not showing
                      major brain functions<br>
                      because they have to deal with the slow
                      oscillations that are very different from the
                      brain's, e.g., as Dr. Henry Markram reported<br>
                      (40Hz?). <br>
                      <br>
                      The above discussion again shows the power and
                      necessity of an overarching brain theory like that
                      in my NAI book. 
                      <br>
                      Those who only simulate biological neurons using
                      superficial biological phenomena are not going to
                      demonstrate
                      <br>
                      any major brain functions.  They can talk about
                      signal statistics from their simulations, but
                      signal statistics are far from brain functions.
                      <br>
                      <br>
                      -John
                      <div>
                        <div class="h5"><br>
                          <br>
                          <div>On 4/8/14 1:30 AM, Tsvi Achler wrote:<br>
                          </div>
                          <blockquote type="cite">
                            <div dir="ltr">Hi John,
                              <div>ART evaluates distance between the
                                contending representation and the
                                current input through vigilance.  If
                                they are too far apart, a poor vigilance
                                signal will be triggered.</div>
                              <div>The best resonance will be achieved
                                when they have the least amount of
                                distance.</div>
                              <div>If in your model, K-nearest neighbors
                                is used without a neural equivalent,
                                then your model is not quite in the
                                spirit of a connectionist model.</div>
                              <div>For example, Bayesian networks do a
                                great job emulating brain behavior,
                                modeling the integration of priors. and
                                has been invaluable to model cognitive
                                studies.  However they assume a
                                statistical configuration of connections
                                and distributions which is not quite
                                known how to emulate with neurons.  Thus
                                pure Bayesian models are also
                                questionable in terms of connectionist
                                modeling.  But some connectionist models
                                can emulate some statistical models for
                                example see section 2.4  in Thomas &
                                McClelland's chapter in Sun's 2008 book
                                (<a moz-do-not-send="true"
href="http://www.psyc.bbk.ac.uk/people/academic/thomas_m/TM_Cambridge_sub.pdf"
                                  target="_blank">http://www.psyc.bbk.ac.uk/people/academic/thomas_m/TM_Cambridge_sub.pdf</a>).</div>
                              <div>I am not suggesting <span
                                  style="font-size:13.333333969116211px;
                                  font-family:arial,sans-serif">Hodgkin-Huxley</span> level
                                detailed neuron models, however
                                connectionist models should have their
                                connections explicitly defined. </div>
                              <div>Sincerely,</div>
                              <div>-Tsvi</div>
                              <div><br>
                              </div>
                            </div>
                            <div class="gmail_extra"><br>
                              <br>
                              <div class="gmail_quote">On Mon, Apr 7,
                                2014 at 10:58 AM, Juyang Weng <span
                                  dir="ltr">
                                  <<a moz-do-not-send="true"
                                    href="mailto:weng@cse.msu.edu"
                                    target="_blank">weng@cse.msu.edu</a>></span>
                                wrote:<br>
                                <blockquote class="gmail_quote"
                                  style="margin:0 0 0 .8ex;
                                  border-left:1px #ccc solid;
                                  padding-left:1ex">
                                  <div bgcolor="#FFFFFF">Tsvi,<br>
                                    <br>
                                    Note that ART uses a vigilance value
                                    to pick up the first "acceptable"
                                    match in its sequential bottom-up
                                    and top-down search.<br>
                                    I believe that was Steve meant when
                                    he mentioned vigilance.    <br>
                                    <br>
                                    Why do you think "ART as a neural
                                    way to implement a K-nearest
                                    neighbor algorithm"? 
                                    <br>
                                    If not all the neighbors have
                                    sequentially participated,<br>
                                    how can ART find the nearest
                                    neighbor, let alone K-nearest
                                    neighbor?<br>
                                    <br>
                                    Our DN uses an explicit k-nearest
                                    mechanism to find the k-nearest
                                    neighbors in every network update,
                                    <br>
                                    to avoid the problems of slow
                                    resonance in existing models of
                                    spiking neuronal networks.  
                                    <br>
                                    The explicit k-nearest mechanism
                                    itself is not meant to be
                                    biologically plausible,
                                    <br>
                                    but it gives a computational
                                    advantage for software simulation of
                                    large networks <br>
                                    at a speed slower than 1000 network
                                    updates per second.<br>
                                    <br>
                                    I guess that more detailed molecular
                                    simulations of individual neuronal
                                    spikes (such as using the
                                    Hodgkin-Huxley model of<br>
                                    a neuron, using the <a
                                      moz-do-not-send="true"
                                      href="http://www.neuron.yale.edu/neuron/"
                                      target="_blank">
                                      NEURON software,</a> or like <a
                                      moz-do-not-send="true"
                                      href="http://bluebrain.epfl.ch/"
                                      target="_blank">
                                      the Blue Brain project</a>
                                    directed by respected Dr. Henry
                                    Markram) <br>
                                    are very useful for showing some
                                    detailed molecular, synaptic, and
                                    neuronal properties.<br>
                                    However, they miss necessary
                                    brain-system-level mechanisms so
                                    much that it is difficult for them
                                    <br>
                                    to show major brain-scale functions
                                    <br>
                                    (such as learning to recognize
                                    objects and detection of natural
                                    objects directly from natural
                                    cluttered scenes).
                                    <br>
                                    <br>
                                    According to my understanding, if
                                    one uses a detailed neuronal model
                                    for each of a variety of neuronal
                                    types and<br>
                                    connects those simulated neurons of
                                    different types according to a
                                    diagram of Brodmann areas,
                                    <br>
                                    his simulation is NOT going to lead
                                    to any major brain function.  <br>
                                    He still needs brain-system-level
                                    knowledge such as that taught in the
                                    BMI 871 course.
                                    <br>
                                    <br>
                                    -John <br>
                                    <div>
                                      <div><br>
                                        <div>On 4/7/14 8:07 AM, Tsvi
                                          Achler wrote:<br>
                                        </div>
                                        <blockquote type="cite">
                                          <div dir="ltr">
                                            <div>Dear Steve, John</div>
                                            I think such discussions are
                                            great to spark interests in
                                            feedback (output back to
                                            input) such models which I
                                            feel should be given much
                                            more attention.
                                            <div>In this vein it may be
                                              better to discuss more of
                                              the details here than to
                                              suggest to read a
                                              reference.</div>
                                            <div><br>
                                            </div>
                                            <div>Basically I see ART as
                                              a neural way to implement
                                              a K-nearest neighbor
                                              algorithm.  Clearly the
                                              way ART overcomes the
                                              neural hurdles is immense
                                              especially in figuring out
                                              how to coordinate neurons.
                                               However it is also
                                              important to summarize
                                              such methods in
                                              algorithmic terms  which I
                                              attempt to do here (and
                                              please comment/correct).</div>
                                            <div>Instar learning is used
                                              to find the best weights
                                              for quick feedforward
                                              recognition without too
                                              much resonance (otherwise
                                              more resonance will be
                                              needed).  Outstar learning
                                              is used to find the
                                              expectation of the
                                              patterns.  The resonance
                                              mechanism evaluates
                                              distances between the
                                              "neighbors" evaluating how
                                              close differing outputs
                                              are to the input pattern
                                              (using the expectation).
                                               By choosing one winner
                                              the network is equivalent
                                              to a 1-nearest neighbor
                                              model.  If you open it up
                                              to more winners (eg k
                                              winners) as you suggest
                                               then it becomes a
                                              k-nearest neighbor
                                              mechanism.</div>
                                            <div><br>
                                            </div>
                                            <div>Clearly I focused here
                                              on the main ART modules
                                              and did not discuss other
                                              additions.  But I want to
                                              just focus on the main
                                              idea at this point.</div>
                                            <div>Sincerely,</div>
                                            <div>-Tsvi</div>
                                          </div>
                                          <div class="gmail_extra"><br>
                                            <br>
                                            <div class="gmail_quote">On
                                              Sun, Apr 6, 2014 at 1:30
                                              PM, Stephen Grossberg <span
                                                dir="ltr">
                                                <<a
                                                  moz-do-not-send="true"
href="mailto:steve@cns.bu.edu" target="_blank">steve@cns.bu.edu</a>></span>
                                              wrote:<br>
                                              <blockquote
                                                class="gmail_quote"
                                                style="margin:0 0 0
                                                .8ex; border-left:1px
                                                #ccc solid;
                                                padding-left:1ex">
                                                <div
                                                  style="word-wrap:break-word"><font
                                                    size="5"
                                                    face="Arial">Dear
                                                    John,</font>
                                                  <div><font size="5"
                                                      face="Arial"><br>
                                                    </font></div>
                                                  <div><font size="5"
                                                      face="Arial">Thanks
                                                      for your
                                                      questions. I reply
                                                      below.</font></div>
                                                  <div><font size="5"
                                                      face="Arial"><br>
                                                    </font>
                                                    <div>
                                                      <div>
                                                        <div><font
                                                          size="5"
                                                          face="Arial">On
                                                          Apr 5, 2014,
                                                          at 10:51 AM,
                                                          Juyang Weng
                                                          wrote:</font></div>
                                                        <font size="5"
                                                          face="Arial"><br>
                                                        </font>
                                                        <blockquote
                                                          type="cite">
                                                          <div
                                                          bgcolor="#FFFFFF"><font
                                                          size="5"
                                                          face="Arial">Dear
                                                          Steve,<br>
                                                          <br>
                                                          This is one of
                                                          my long-time
                                                          questions that
                                                          I did not have
                                                          a chance to
                                                          ask you when I
                                                          met you many
                                                          times before.
                                                          <br>
                                                          But they may
                                                          be useful for
                                                          some people on
                                                          this list.   <br>
                                                          Please accept
                                                          my apology of
                                                          my question
                                                          implies any
                                                          false
                                                          impression
                                                          that I did not
                                                          intend.<br>
                                                          <br>
                                                          (1) Your
                                                          statement
                                                          below seems to
                                                          have confirmed
                                                          my
                                                          understanding: 
                                                          <br>
                                                          Your top-down
                                                          process in ART
                                                          in the late
                                                          1990's is
                                                          basically for
                                                          finding an
                                                          acceptable
                                                          match
                                                          <br>
                                                          between the
                                                          input feature
                                                          vector and the
                                                          stored feature
                                                          vectors
                                                          represented by
                                                          neurons (not
                                                          meant for the
                                                          nearest
                                                          match).
                                                          <br>
                                                          </font></div>
                                                        </blockquote>
                                                        <div><font
                                                          size="5"
                                                          face="Arial"><br>
                                                          </font></div>
                                                      </div>
                                                      <font size="5"
                                                        face="Arial">ART
                                                        has developed a
                                                        lot since the
                                                        1990s. A
                                                        non-technical
                                                        but fairly
                                                        comprehensive
                                                        review article
                                                        was published in
                                                        2012 in
                                                        <i>Neural
                                                          Networks</i>
                                                        and can be found
                                                        at <a
                                                          moz-do-not-send="true"
href="http://cns.bu.edu/%7Esteve/ART.pdf" target="_blank">http://cns.bu.edu/~steve/ART.pdf</a>.</font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">I
                                                        do not think
                                                        about the
                                                        top-down process
                                                        in ART in quite
                                                        the way that you
                                                        state above. My
                                                        reason for this
                                                        is summarized by
                                                        the acronym
                                                        CLEARS for the
                                                        processes of
                                                        Consciousness,
                                                        Learning,
                                                        Expectation,
                                                        Attention,
                                                        Resonance, and
                                                        Synchrony. </font><span
                                                        style="font-family:Arial;
font-size:x-large">All the CLEARS processes come into this story, and </span><span
                                                        style="font-family:Arial;
font-size:x-large">ART top-down mechanisms contribute to all of them.
                                                        For me, the most
                                                        fundamental
                                                        issues concern
                                                        how ART
                                                        dynamically
                                                        self-stabilizes
                                                        the memories
                                                        that are learned
                                                        within the
                                                        model's
                                                        bottom-up
                                                        adaptive filters
                                                        and top-down
                                                        expectations. </span></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">In
                                                        particular,
                                                        during learning,
                                                        a big enough
                                                        mismatch can
                                                        lead to
                                                        hypothesis
                                                        testing and
                                                        search for a
                                                        new, or
                                                        previously
                                                        learned,
                                                        category that
                                                        leads to an
                                                        acceptable
                                                        match. The
                                                        criterion for
                                                        what is "big
                                                        enough mismatch"
                                                        or "acceptable
                                                        match" is
                                                        regulated by a
                                                        vigilance
                                                        parameter that
                                                        can itself vary
                                                        in a
                                                        state-dependent
                                                        way.</font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">After
                                                        learning occurs,
                                                        a bottom-up
                                                        input pattern
                                                        typically
                                                        directly selects
                                                        the
                                                        best-matching
                                                        category,
                                                        without any
                                                        hypothesis
                                                        testing or
                                                        search. And even
                                                        if there is a
                                                        reset due to a
                                                        large initial
                                                        mismatch with a
                                                        previously
                                                        active category,
                                                        a single reset
                                                        event may lead
                                                        directly to a
                                                        matching
                                                        category that
                                                        can directly
                                                        resonate with
                                                        the data. </font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">I
                                                        should note that
                                                        all of the
                                                        foundational
                                                        predictions of
                                                        ART now have
                                                        substantial
                                                        bodies of
                                                        psychological
                                                        and
                                                        neurobiological
                                                        data to support
                                                        them. See the
                                                        review article
                                                        if you would
                                                        like to read
                                                        about them.</font></div>
                                                    <div>
                                                      <div><font
                                                          size="5"
                                                          face="Arial"><br>
                                                        </font>
                                                        <blockquote
                                                          type="cite">
                                                          <div
                                                          bgcolor="#FFFFFF"><font
                                                          size="5"
                                                          face="Arial">The
                                                          currently
                                                          active neuron
                                                          is the one
                                                          being examined
                                                          by the top
                                                          down process<br>
                                                          </font></div>
                                                        </blockquote>
                                                        <div><font
                                                          size="5"
                                                          face="Arial"><br>
                                                          </font></div>
                                                      </div>
                                                      <font size="5"
                                                        face="Arial">I'm
                                                        not sure what
                                                        you mean by
                                                        "being
                                                        examined", but
                                                        perhaps my
                                                        comment above
                                                        may deal with
                                                        it.</font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">I
                                                        should comment,
                                                        though, about
                                                        your use of the
                                                        word "currently
                                                        active neuron".
                                                        I assume that
                                                        you mean at the
                                                        category level. </font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">In
                                                        this
                                                        regard, there
                                                        are two ART's.
                                                        The first aspect
                                                        of ART is as a
                                                        cognitive and
                                                        neural theory
                                                        whose scope,
                                                        which includes
                                                        perceptual,
                                                        cognitive, and
                                                        adaptively timed
                                                        cognitive-emotional
                                                        dynamics, among
                                                        other processes,
                                                        is illustrated
                                                        by the above
                                                        referenced 2012
                                                        review article
                                                        in <i>Neural
                                                          Networks</i>.
                                                        In the
                                                        biological
                                                        theory, there is
                                                        no general
                                                        commitment to
                                                        just one
                                                        "currently
                                                        active neuron".
                                                        One always
                                                        considers the
                                                        neuronal
                                                        population, or
                                                        populations,
                                                        that represent a
                                                        learned
                                                        category.
                                                        Sometimes, but
                                                        not always, a
                                                        winner-take-all
                                                        category is
                                                        chosen. </font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">The
                                                        2012 review
                                                        article
                                                        illustrates some
                                                        of the large
                                                        data bases of
                                                        psychological
                                                        and
                                                        neurobiological
                                                        data that have
                                                        been explained
                                                        in a principled
                                                        way,
                                                        quantitatively
                                                        simulated, and
                                                        successfully
                                                        predicted by ART
                                                        over a period of
                                                        decades.
                                                        ART-like
                                                        processing is,
                                                        however,
                                                        certainly not
                                                        the only kind of
                                                        computation that
                                                        may be needed to
                                                        understand how
                                                        the brain works.
                                                        The paradigm
                                                        called
                                                        Complementary
                                                        Computing that I
                                                        introduced
                                                        awhile ago makes
                                                        precise the
                                                        sense in which
                                                        ART may be just
                                                        one kind of
                                                        dynamics
                                                        supported by
                                                        advanced brains.
                                                        This is also
                                                        summarized in
                                                        the review
                                                        article.<br>
                                                      </font>
                                                      <div><font
                                                          size="5"
                                                          face="Arial"><br>
                                                        </font></div>
                                                      <div><font
                                                          size="5"
                                                          face="Arial">The
                                                          second aspect
                                                          of ART is as a
                                                          series of
                                                          algorithms
                                                          that
                                                          mathematically
                                                          characterize
                                                          key ART design
                                                          principles and
                                                          mechanisms in
                                                          a focused
                                                          setting, and
                                                          provide
                                                          algorithms for
                                                          large-scale
                                                          applications
                                                          in engineering
                                                          and
                                                          technology.
                                                          ARTMAP, fuzzy
                                                          ARTMAP, and
                                                          distributed
                                                          ARTMAP are
                                                          among these,
                                                          all of them
                                                          developed with
                                                          Gail
                                                          Carpenter.
                                                          Some of these
                                                          algorithms use
                                                          winner-take-all
                                                          categories to
                                                          enable the
                                                          proof of
                                                          mathematical
                                                          theorems that
                                                          characterize
                                                          how underlying
                                                          design
                                                          principles
                                                          work. In
                                                          contrast, the
                                                          distributed
                                                          ARTMAP family
                                                          of algorithms,
                                                          developed by
                                                          Gail Carpenter
                                                          and her
                                                          colleagues,
                                                          allows for
                                                          distributed
                                                          category
                                                          representations
                                                          without losing
                                                          the benefits
                                                          of fast,
                                                          incremental,
                                                          self-stabilizing
                                                          learning and
                                                          prediction in
                                                          response to a
                                                          large
                                                          non-stationary
                                                          databases that
                                                          can include
                                                          many
                                                          unexpected
                                                          events. </font></div>
                                                      <div><font
                                                          size="5"
                                                          face="Arial"><br>
                                                        </font></div>
                                                      <div><font
                                                          size="5"
                                                          face="Arial">See,
                                                          e.g., <a
                                                          moz-do-not-send="true"
href="http://techlab.bu.edu/members/gail/articles/115_dART_NN_1997_.pdf"
target="_blank">http://techlab.bu.edu/members/gail/articles/115_dART_NN_1997_.pdf</a>
                                                          and
                                                          <a
                                                          moz-do-not-send="true"
href="http://techlab.bu.edu/members/gail/articles/155_Fusion2008_CarpenterRavindran.pdf"
target="_blank">
http://techlab.bu.edu/members/gail/articles/155_Fusion2008_CarpenterRavindran.pdf</a>.</font></div>
                                                      <div><font
                                                          size="5"
                                                          face="Arial"><br>
                                                        </font></div>
                                                      <div><font
                                                          size="5"
                                                          face="Arial">I
                                                          should note
                                                          that FAST
                                                          learning is a
                                                          technical
                                                          concept: it
                                                          means that
                                                          each adaptive
                                                          weight can
                                                          converge to
                                                          its new
                                                          equilibrium
                                                          value on EACH
                                                          learning
                                                          trial. That is
                                                          why ART
                                                          algorithms can
                                                          often
                                                          successfully
                                                          carry out
                                                          one-trial
                                                          incremental
                                                          learning of a
                                                          data base.
                                                          This is not
                                                          true of many
                                                          other
                                                          algorithms,
                                                          such as back
                                                          propagation,
                                                          simulated
                                                          annealing, and
                                                          the like,
                                                          which all
                                                          experience
                                                          catastrophic
                                                          forgetting if
                                                          they try to do
                                                          fast learning.
                                                          Almost all
                                                          other learning
                                                          algorithms
                                                          need to be run
                                                          using slow
                                                          learning, that
                                                          allows only a
                                                          small
                                                          increment in
                                                          the values of
                                                          adaptive
                                                          weights on
                                                          each learning
                                                          trial, to
                                                          avoid massive
                                                          memory
                                                          instabilities,
                                                          and work best
                                                          in response to
                                                          stationary
                                                          data. Such
                                                          algorithms
                                                          often fail to
                                                          detect
                                                          important rare
                                                          cases, among
                                                          other
                                                          limitations.
                                                          ART can
                                                          provably learn
                                                          in either the
                                                          fast or slow
                                                          mode in
                                                          response to
                                                          non-stationary
                                                          data.</font></div>
                                                      <div>
                                                        <div><font
                                                          size="5"
                                                          face="Arial"><br>
                                                          </font></div>
                                                        <blockquote
                                                          type="cite">
                                                          <div
                                                          bgcolor="#FFFFFF"><font
                                                          size="5"
                                                          face="Arial">in
                                                          a sequential
                                                          fashion: one
                                                          neuron after
                                                          another, until
                                                          an acceptable
                                                          neuron is
                                                          found.<br>
                                                          <br>
                                                          (2) The input
                                                          to the ART in
                                                          the late
                                                          1990's is for
                                                          a single
                                                          feature vector
                                                          as a
                                                          monolithic
                                                          input. 
                                                          <br>
                                                          By monolithic,
                                                          I mean that
                                                          all neurons
                                                          take the
                                                          entire input
                                                          feature vector
                                                          as input.  
                                                          <br>
                                                          I raise this
                                                          point here
                                                          because neuron
                                                          in ART in the
                                                          late 1990's
                                                          does not have
                                                          an explicit
                                                          local sensory
                                                          receptive
                                                          field (SRF),
                                                          <br>
                                                          i.e., are
                                                          fully
                                                          connected from
                                                          all components
                                                          of the input
                                                          vector.   A
                                                          local SRF
                                                          means that
                                                          each neuron is
                                                          only connected
                                                          to a small
                                                          region
                                                          <br>
                                                          in an input
                                                          image. <br>
                                                          </font></div>
                                                        </blockquote>
                                                        <div><font
                                                          size="5"
                                                          face="Arial"><br>
                                                          </font></div>
                                                      </div>
                                                      <font size="5"
                                                        face="Arial">Various
                                                        ART algorithms
                                                        for technology
                                                        do use fully
                                                        connected
                                                        networks. They
                                                        represent a
                                                        single-channel
                                                        case, which is
                                                        often sufficient
                                                        in applications
                                                        and which
                                                        simplifies
                                                        mathematical
                                                        proofs. However,
                                                        the
                                                        single-channel
                                                        case is, as its
                                                        name suggests,
                                                        not a necessary
                                                        constraint on
                                                        ART design. </font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">In
                                                        addition, many
                                                        ART biological
                                                        models do not
                                                        restrict
                                                        themselves to
                                                        the
                                                        single-channel
                                                        case, and do
                                                        have receptive
                                                        fields. These
                                                        include the
                                                        LAMINART family
                                                        of models that
                                                        predict
                                                        functional roles
                                                        for many
                                                        identified cell
                                                        types in the
                                                        laminar circuits
                                                        of cerebral
                                                        cortex. These
                                                        models
                                                        illustrate how
                                                        variations of a
                                                        shared laminar
                                                        circuit design
                                                        can carry out
                                                        very different
                                                        intelligent
                                                        functions, such
                                                        as 3D vision
                                                        (e.g., 3D
                                                        LAMINART),
                                                        speech and
                                                        language (e.g.,
                                                        cARTWORD), and
                                                        cognitive
                                                        information
                                                        processing
                                                        (e.g., LIST
                                                        PARSE). They are
                                                        all summarized
                                                        in the 2012
                                                        review article,
                                                        with the
                                                        archival
                                                        articles
                                                        themselves on my
                                                        web page
                                                        <a
                                                          moz-do-not-send="true"
href="http://cns.bu.edu/%7Esteve" target="_blank">http://cns.bu.edu/~steve</a>. </font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">The
                                                        existence of
                                                        these laminar
                                                        variations-on-a-theme
                                                        provides an
                                                        existence proof
                                                        for the exciting
                                                        goal of
                                                        designing a
                                                        family of chips
                                                        whose
                                                        specializations
                                                        can realize all
                                                        aspects of
                                                        higher
                                                        intelligence,
                                                        and which can be
                                                        consistently
                                                        connected
                                                        because they all
                                                        share a similar
                                                        underlying
                                                        design. Work on
                                                        achieving this
                                                        goal can
                                                        productively
                                                        occupy lots of
                                                        creative
                                                        modelers and
                                                        technologists
                                                        for many years
                                                        to come.</font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">I
                                                        hope that the
                                                        above replies
                                                        provide some
                                                        relevant
                                                        information, as
                                                        well as pointers
                                                        for finding
                                                        more.</font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">Best,</font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial">Steve</font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font></div>
                                                    <div>
                                                      <blockquote
                                                        type="cite">
                                                        <div
                                                          bgcolor="#FFFFFF">
                                                          <div><font
                                                          size="5"
                                                          face="Arial"><br>
                                                          My apology
                                                          again if my
                                                          understanding
                                                          above has
                                                          errors
                                                          although I
                                                          have examined
                                                          the above two
                                                          points
                                                          carefully
                                                          <br>
                                                          through
                                                          multiple your
                                                          papers.<br>
                                                          <br>
                                                          Best regards,<br>
                                                          <br>
                                                          -John<br>
                                                          <br>
                                                          </font></div>
                                                          <div>
                                                          <pre cols="72"><font face="Arial"><span style="font-size:18px">Juyang (John) Weng, Professor
Department of Computer Science and Engineering
MSU Cognitive Science Program and MSU Neuroscience Program
428 S Shaw Ln Rm 3115
Michigan State University
East Lansing, MI 48824 USA
Tel: <a moz-do-not-send="true" href="tel:517-353-4388" value="+15173534388" target="_blank">517-353-4388</a>
Fax: <a moz-do-not-send="true" href="tel:517-432-1061" value="+15174321061" target="_blank">517-432-1061</a>
Email: <a moz-do-not-send="true" href="mailto:weng@cse.msu.edu" target="_blank">weng@cse.msu.edu</a>
URL: <a moz-do-not-send="true" href="http://www.cse.msu.edu/%7Eweng/" target="_blank">http://www.cse.msu.edu/~weng/</a>
----------------------------------------------

</span></font></pre>
                                                          </div>
                                                        </div>
                                                      </blockquote>
                                                    </div>
                                                    <div><font size="5"
                                                        face="Arial"><br>
                                                      </font>
                                                      <div><font
                                                          size="5"
                                                          face="Arial"><span
                                                          style="line-height:normal;
                                                          text-indent:0px;
                                                          border-collapse:separate;
                                                          letter-spacing:normal;
                                                          font-variant:normal;
                                                          text-transform:none;
                                                          font-style:normal;
                                                          white-space:normal;
                                                          font-weight:normal;
word-spacing:0px"><span style="line-height:normal; text-indent:0px;
                                                          border-collapse:separate;
                                                          letter-spacing:normal;
                                                          font-variant:normal;
                                                          text-transform:none;
                                                          font-style:normal;
                                                          white-space:normal;
                                                          font-weight:normal;
word-spacing:0px">
                                                          <div
                                                          style="word-wrap:break-word"><span
                                                          style="line-height:normal;
                                                          text-indent:0px;
                                                          border-collapse:separate;
                                                          letter-spacing:normal;
                                                          font-variant:normal;
                                                          text-transform:none;
                                                          font-style:normal;
                                                          white-space:normal;
                                                          font-weight:normal;
word-spacing:0px">
                                                          <div
                                                          style="word-wrap:break-word">
                                                          <div>
                                                          <div>
                                                          <div>
                                                          <div>Stephen
                                                          Grossberg</div>
                                                          <div>Wang
                                                          Professor of
                                                          Cognitive and
                                                          Neural Systems</div>
                                                          <div>Professor
                                                          of
                                                          Mathematics,
                                                          Psychology,
                                                          and Biomedical
                                                          Engineering</div>
                                                          <div>
                                                          <div>Director,
                                                          Center for
                                                          Adaptive
                                                          Systems <a
                                                          moz-do-not-send="true"
href="http://www.cns.bu.edu/about/cas.html" target="_blank">http://www.cns.bu.edu/about/cas.html</a></div>
                                                          </div>
                                                          <div><a
                                                          moz-do-not-send="true"
href="http://cns.bu.edu/%7Esteve" target="_blank">http://cns.bu.edu/~steve</a></div>
                                                          <div><a
                                                          moz-do-not-send="true"
href="mailto:steve@bu.edu" target="_blank">steve@bu.edu</a></div>
                                                          </div>
                                                          </div>
                                                          </div>
                                                          <div><br>
                                                          </div>
                                                          </div>
                                                          </span></div>
                                                          </span><br>
                                                          </span><br>
                                                        </font></div>
                                                      <br>
                                                    </div>
                                                  </div>
                                                </div>
                                              </blockquote>
                                            </div>
                                            <br>
                                          </div>
                                        </blockquote>
                                        <br>
                                      </div>
                                    </div>
                                    <pre cols="72"><span><font color="#888888">-- 
--
Juyang (John) Weng, Professor
Department of Computer Science and Engineering
MSU Cognitive Science Program and MSU Neuroscience Program
428 S Shaw Ln Rm 3115
Michigan State University
East Lansing, MI 48824 USA
Tel: <a moz-do-not-send="true" href="tel:517-353-4388" value="+15173534388" target="_blank">517-353-4388</a></font></span><div>
Fax: <a moz-do-not-send="true" href="tel:517-432-1061" value="+15174321061" target="_blank">517-432-1061</a>
Email: <a moz-do-not-send="true" href="mailto:weng@cse.msu.edu" target="_blank">weng@cse.msu.edu</a>
URL: <a moz-do-not-send="true" href="http://www.cse.msu.edu/%7Eweng/" target="_blank">http://www.cse.msu.edu/~weng/</a>
----------------------------------------------

</div></pre>
                                  </div>
                                </blockquote>
                              </div>
                              <br>
                            </div>
                          </blockquote>
                          <br>
                          <pre cols="72">-- 
--
Juyang (John) Weng, Professor
Department of Computer Science and Engineering
MSU Cognitive Science Program and MSU Neuroscience Program
428 S Shaw Ln Rm 3115
Michigan State University
East Lansing, MI 48824 USA
Tel: <a moz-do-not-send="true" href="tel:517-353-4388" value="+15173534388" target="_blank">517-353-4388</a>
Fax: <a moz-do-not-send="true" href="tel:517-432-1061" value="+15174321061" target="_blank">517-432-1061</a>
Email: <a moz-do-not-send="true" href="mailto:weng@cse.msu.edu" target="_blank">weng@cse.msu.edu</a>
URL: <a moz-do-not-send="true" href="http://www.cse.msu.edu/%7Eweng/" target="_blank">http://www.cse.msu.edu/~weng/</a>
----------------------------------------------

</pre>
                        </div>
                      </div>
                    </div>
                  </blockquote>
                </div>
                <br>
              </div>
            </blockquote>
            <br>
            <br>
            <pre class="moz-signature" cols="72">-- 
Prof. Dr. Danko Nikolić

Web:
<a moz-do-not-send="true" class="moz-txt-link-freetext" href="http://www.danko-nikolic.com">http://www.danko-nikolic.com</a>

Mail address 1:
Department of Neurophysiology
Max Planck Institut for Brain Research
Deutschordenstr. 46
60528 Frankfurt am Main
GERMANY

Mail address 2:
Frankfurt Institute for Advanced Studies
Wolfgang Goethe University
Ruth-Moufang-Str. 1
60433 Frankfurt am Main
GERMANY

----------------------------
Office: (..49-69) 96769-736
Lab: (..49-69) 96769-209
Fax: (..49-69) 96769-327
<a moz-do-not-send="true" class="moz-txt-link-abbreviated" href="mailto:danko.nikolic@gmail.com">danko.nikolic@gmail.com</a>
----------------------------</pre>
          </div>
        </div>
      </div>
    </blockquote>
    <br>
    <br>
    <pre class="moz-signature" cols="72">-- 
Prof. Dr. Danko Nikolić

Web:
<a class="moz-txt-link-freetext" href="http://www.danko-nikolic.com">http://www.danko-nikolic.com</a>

Mail address 1:
Department of Neurophysiology
Max Planck Institut for Brain Research
Deutschordenstr. 46
60528 Frankfurt am Main
GERMANY

Mail address 2:
Frankfurt Institute for Advanced Studies
Wolfgang Goethe University
Ruth-Moufang-Str. 1
60433 Frankfurt am Main
GERMANY

----------------------------
Office: (..49-69) 96769-736
Lab: (..49-69) 96769-209
Fax: (..49-69) 96769-327
<a class="moz-txt-link-abbreviated" href="mailto:danko.nikolic@gmail.com">danko.nikolic@gmail.com</a>
----------------------------</pre>
  </body>
</html>