[Comp-neuro] High level abstractions in concept cells; Single spiking neurons have "meaning" and are actually at the cognitive level?

Dorian Aur dorianaur at gmail.com
Thu Aug 11 21:54:10 CEST 2011

Asim, your idea to add together various scientists with similar interests
represents an important step to trigger a debate that can  indeed solve
current issues regarding data analysis and the nature of neural code.

I’ve had myself the privilege to analyze some  excellent recordings from Dr
Fried lab. In general it is an exception to get four good neurons where
spike directivity (SD) can be  computed (the electrodes need to be together
in a tetrode configuration). Dr Fried I’m really impressed, excellent
recordings! Therefore  in an attempt of  *pursuing the truth *I have
analyzed some of these recordings from a different perspective *and *took
the liberty to *post it here the outcome.*

*Dorian’s summary of some issues in interpreting temporal coding - includes
notes from **Dr. Itzhak Fried and Christof Koch, **Quian Quiroga, **Walter

*1.*       Work on brain-machine interfaces based on single-neuron activity
is quite standard now and being pursued by a number of groups.

*Itzhak Fried: **I would not describe BMI using single neuron activity
as **quite
standard. Most of the existing data is with frontal/motor or parietal

*Christof:** **Yes. Most such BMI operate in motor, pre-motor and parietal
cortex. *

*Dorian :*There are several issues in interpreting single neuron activity
using the  firing rate or interspike interval. Several  important details
are  missing  in temporal patterns (see
http://dx.doi.org./10.1016/j.jneumeth.2005.05.006  and  the book
Neuroelectrodynamics: Understanding the Brain Language,
http://dx.doi.org/10.3233/978-1-60750-473-3-i for a model of computation)
which are important in information coding

*2.*       Concept cells are comparable to place cells in rodents (concept
cells = place cells) and therefore not a finding that surprises the
neuroscience community.

*Itzhak Fried: **Concept cells are not place cells but I proposed that they
can be viewed as "place cells" in a different "attribute or feature space".
They do share with place cells coding properties, that is: specificity ,
invariance, sparseness and the explicit nature of the code. One can
speculate that the mechanism developed for coding of space in rodent
hippocampus  has evolved to accommodate more elaborate abstraction in
humans. As for "surprises", it is difficult to surprise the neuroscience
community, but for us the explicit nature of the code on the single neuron
level was a surprise.*

*Christof:** **There are some similarities to place cells in rodents.
However, we find these highly selective cells in all regions of the MTL, not
just the hippocampus.  How far this comparison goes is not clear (what, for
example, is the analog of grid cells in the entorhinal cortex?)*

*Dorian:** *Different experiments in rats or humans show strong similarities
that can provide meaningful explanation for these data. However, an
understanding of presented examples cannot solely come from a firing rate
analysis. A relevant example shows how neurons operate during  a T-maze
procedural learning task. (D Aur, and M Jog,  Reading the Neural Code: What
do Spikes Mean for Behavior?. Available from Nature Precedings <
http://dx.doi.org/10.1038/npre.2007.61.1, 2007)

The "expert" neurons in striatum during  T-maze learning provide a similar
behavior,  however *they  extensively fire only before learning*. In order
to understand the meaning of their firing activity a different measure was
computed and analyzed. *Spike directivity* is a vector that reflects the
distribution of electrical patterns in recorded spikes. During learning
these  "expert" neurons reduce their firing rate. After one week of training
the neurons generate only few spikes between the tone and  turn starts.  This
represents the critical moment when the decision regarding turning is taken.
After training these “expert” neurons show less random  spike directivities
 (a preferred direction of AP propagation)than before training . The
delivered spikes after  the tone  predict the turning direction on the
T-maze. In many cases the firing rate cannot be estimated  (one spike in
single trials)
*The first counterexample: *When it fires the same neuron can code for the
left turn or for the right turn depending on the context (high or low  tone,
the T-maze task,
http://neuroelectrodynamics.blogspot.com/p/spike-directivity.html ).

*The second counterexample:* The same neuron responds with the same firing
rate for two different  objects (spider, Jennifer Aniston) the difference
occurs  in the preferred spike directivity (see

The outcome in spike directivity is a counterexample for temporal coding. *The
spikes cannot be added since they provide different semantics (apples and
oranges). That’s the beauty of counterexamples. You only need one single
counterexample to throw down a "solid"theoretical construct  of temporal
coding. No need  for other examples that reinforce the  temporal coding!

In this case:
(i) During learning several options are explored and the  *strong firing
rate reflects uncertainty in these “expert” neurons.  Therefore, we
hypothesized that strong firing (with strong variability of spike
directivity) represents a way * to search for a  correct solution during
(ii) After learning all these cells provide an efficient response with only
few spikes for the same event.

(iii) The *reduction in uncertainty generates a meaningful outcome* that can
be observed in a preferred direction of spike directivity
Therefore, *a decrease of uncertainty  is reflected in a reduced number of
spikes delivered by a cell, an efficient response*. Contrary to current
belief an increase in the firing rate *may show uncertainty*,  a  searching
 process required to deliver a solution.

Following  a similar analysis,  the cells from MTL can display a similar
behavior. If two different objects are presented they can be separated  fast
in these neurons  since they generate different spike
directivity orientations (see

*3.*       The concept cells were found in different regions. For example,
“James Brolin” in right hippocampus, “Venus Williams” in left hippocampus,
“Marilyn Monroe” in left parahippocampal cortex, “Michael Jackson” in right

*Itzhak Fried: **Yes, but they may represent different levels of abstraction
or invariance in each of these regions. Although they were found in
different MTL regions , the highest degree of invariance (across modalities)
was in hippocampus and entorhinal cortex . Also remember the latency of the
response,  usually around 350 msec.*

*Christof:** **Yes*

*Dorian :I*ntracellularly  within molecular structure specific information
is "read" and "written" in these cells  during AP generation . In order to
have the concept of "Jennifer Aniston" information from many cells is
inferred synaptically and non-synaptically. Therefore, many neurons fire
almost simultaneously in different brain regions  (MTL hippocampus ,
entorhinal cortex)

(i)                 However, a  too strong  increase of firing rate may  show
high uncertainty - a searching process required to identify the
presented  object.

(ii)               In order to represent a particular feature associated
with  a certain presented image (e.g. Jennifer Aniston) these neuron can
generate low firing rate with consistent  preferred spike directivity

(iii)             The semantics do not appear  in the firing rate!!! (ISI)
therefore statistical analyzes  of firing rate (ISI)  can be highly
irrelevant to determine the meaning of firing. The same neuron can code for
different features in different spikes depending on presented context.

*4.*       The sister cells (e.g. other Jennifer Aniston concept cells) are
not necessarily in contiguous locations in the brain. They could be in
different hemispheres and different regions within a hemisphere. (“The
subject most likely activated a large pool of neurons selective to ‘Johnny
Cash’ even though the feedback was only based on just one such unit. We
identified 8 such units in a total of 7 subjects.”)

*Itzhak Fried: **The sister cell may be  a confusing term, but a major point
is that organization of "concept cells"  is not columnar or topographic.
Given their sparse and nontopographic distribution  it would be difficult to
trace them on fMRI.*

*Dorian: *There is little information in the temporal code (firing rate,
Based solely on firing rate (ISI) analyses it is very hard to figure out  the
role of certain spikes ( see the counterexample where the same neuron
provides different semantics when it fires
http://precedings.nature.com/documents/61/version/1) .* ***

*5.*       Even though a million cells are activated by an image of Jennifer
Aniston, and say 12 of them are Jennifer Aniston concept cells, in your
experiments, you tracked only one such concept cell and* **that was good
enough*.* **There was no need to “read out” other Jennifer Aniston concept
cells*, wherever they were,* **as would be required in a distributed
representation framework*.

*Itzhak Fried: **Yes. But I  suspect more than a million cells are activated
by Jennifer Aniston and they could probably be arranged on a variance scale
with our "concept cells" at the extreme low. Still it is easier to find a
concept cell than a Higg's boson.*

*Christof:** **We have no idea whether J. Aniston activates a million of
such cells. Yes, the movie of the superimposed images was based on four
selective units of a presumably much larger pool. It is well possible that
if we had recorded from more sister neurons, control would have been swifter
or more precise or more reliable.*

*Dorian: *

(i)These cells do not fire only for Jennifer Aniston as presented in Quiroga
et al., 2005. (ii) In different contexts they should   fire for different
presented objects (see http://precedings.nature.com/documents/5345/version/2).

(iii) A strong  increase  in the firing rate  may show a different process
(iv)If spike directivity points randomly in space then the  strong increase
in firing rate display uncertainty, an ongoing  “searching” process to
associate different presented features
(v)The efficient coding of a particular feature associated with Jennifer
Aniston  needs to provide a consistent preferred spike directivity. This
outcome is determined by a  consistent intracellular location of particular
"memories" and is revealed using spike directivity or imaging the spike.
(see http://precedings.nature.com/documents/5345/version/2 or

6.       In your image control experiments, where the subject focused on one
of two images on a computer screen to enhance its visibility (a target vs. a
distractor image)* **by “thinking” about it* (the target image),* **the
subjects were able to control and modulate the activity of the concept units
selective to specific images*. “Thinking” in this case might simply imply
invoking some images from memory of the target concept (e.g. Jennifer
Aniston) and that might also imply the “internal assignment” of meaning to
the target concept cells. (This is a tenuous argument. Wish we could also
say that the concept cell activates the “memories,” thereby providing
linkage both ways.)

*Itzhak Fried:  **Read the Science 2008 paper by Gelbard -Sagiv, Mukamel,
Harel, Malach and Fried where you will see how such cell (firing selectively
to the actual sight of a 10 sec video of the Simpson's, as one example ) is
reactivated just before (1.2 sec)  the patient reports the recollection.*

*Christof:** **The cognitive or neuronal processes underlying the voluntary
control seen in these fading experiments are unclear. Personally, I think
they are closer to object based attention than to memory but this remains to
be **proven.*

*Dorian: *I was particularly interested by  the *Simpson *example,  even
stopped the presentation to show that  this specific cell  has fired  for
different other images with low firing rate. If during the increase in
firing rate the computed  spike directivities show  outcome then this case
can be a typical example where the neuron  is “searching” for a solution.
Since  other* different  neurons may activate the recollection* *the
process can be triggered.  I feel that this neuron will not provide  stable
high firing rate longer time in this case (*over one week of repeated
Simpson presentation*).    **If spike directivity is less  random then
indeed this particular neuron can embed some  features  associated to
Simpson. Information  is “read” or “written” in this cell during these
spikes and  the cell  may contribute to form the Simpson abstraction(
however not alone!).***

7.       Here’s an interesting conclusion from Waydo, Kraskov, Quiroga,
Fried and Koch (The Journal of Neuroscience, 2006):

“Instead, it would imply that rather than a single neuron responding to
dozens of stimuli out of a universe of tens of thousands,* **such a neuron
might respond to only one or a few stimuli out of perhaps hundreds currently
being tracked by this memory system*, still with millions of neurons being
activated by a typical stimulus. These results are consistent with Barlow’s
(1972) claim that “at the upper levels of the hierarchy,* **a relatively
small proportion [of neurons] are active, and each of these says a lot when
it is active*,” and his further speculation that the “*aim of information
processing in higher sensory centers is to represent the input as completely
as possible by activity in as few neurons as possible*” (Barlow, 1972).”

*Itzhak Fried: **When I proposed the term "concept cells" for the unique
group of cells we found in hippocampus and neighbouring MTL structures it
was with the intention of provoking such diuscussions, but using the
nomenclature we should not be carried away by the hype of the terminology
and lose sight of the data.(I do agree with Freeman's cautionary note re
"meaning"). Do not forget that these cells are at the heart of the
declarative memory system of MTL and thus  signify the transformation of
percepts into what can be later consciously recollected.*

*The intriguing question is how these cells are formed and change. We know
patients form these cells to the experimenters over a day or so.  We are
currently completing  a study which will provide some relevant data.*

*Christof:** **Yes, Horace Barlow's 1972 paper was very forward looking and
deserves to be widely read and cited. *The efficiency in processing
information is the main goal of the brain, therefore the process of  object
recognition is optimized in these cells that respond  which a decrease in
firing rate and a “specialization” of involved neurons that carries specific

*Quian Quiroga*

Asim, thanks for triggering this interesting discussion.

Yes, I do believe these cells encode meaning. We say this explicitly in a
TiCS paper (at the end of the section before the Conclusion).

*From Walter Freeman*

Your paraphrase is ambiguous. "Concept cells" certainly have meaning for
observers, but do they express and transmit meaning within the brain of the
subject to other parts of the brain? In other words, how in a small fraction
of a second does the output of the "concept cell" capture and control
attention and the neural machinery leading to Sherrington's "final common

I conceive the "concept cell" as one of ~10^5 neurons forming a Hebbian
assembly, which provides the key to a global attractor and the energy needed
to  trigger a phase transition. In this view the meaning is expressed by the
attractor involving ~10^9 neurons. The spikes of the sampled "concept cell"
(in concert with ~10^5 - 1 other cells) are an essential sign, neural
correlate, and agency mediating the construction of meaning from the memory
(synaptic matrix) selected by a stimu

*Dorian: *The efficiency of information processing  seems to be  the main
reason of changes in the dynamics of firing. The T-maze learning shows a
process of optimization. The firing rate is reduced when  a certain
semantics is acquired in single cells. Here, in these recordings  it seem to
be a similar process. If the objects are presented several times the
increase in  “specialization”  occurs

I agree with Dr Freeman.  The temporal coding is ambiguous (see both
counterexamples) the meaning  seems to be a result of electrical inference (not
of temporal patterns) http://precedings.nature.com/documents/5345/version/2)
- .

The *Horace Barlow's 1972 paper was an inspiration to  develop the new
model – NeuroElectroDynamics (NED). This computational model shows that
information is integrated across different  scales  in the brain using
electrical activity  (not temporal patterns) in order to generate the *“the
*final common path” (* see a small network of four neurons
http://precedings.nature.com/documents/5345/version/2) *. *Many scientists
have previously envisioned and described different (non-Turing) forms of
computations. *Computing by physical interaction in neurons ( in the brain)
generates a powerful (non-Turing)  model of computation.

I’m  always  interested to analyze good recordings and really  delighted
that  with these  new techniques the mystery of neural code can be solved
the neural code was not the main goal goal. The result occurred in response
to  other different questions- Why artificial intelligence cannot move
beyond  capabilities of a two-and-a-half year old child? Why brain
computations are so powerful?

We found that several controversies in the field were generated by keeping
alive the temporal coding paradigm. I value the contribution of all
scientists that have worked in this field; they kept our interest focused on
fundamental issues and our success in understanding how the brain computes
(see NED) reflects  their long-standing  effort in this area.

*"We have seen a little further by standing on the shoulders of Giants  ....
not because our sight is superior or because we are taller than they, but
because they raise us up, and by their great stature add to ours”*

*Isaac Newton*

Therefore, I'm actively interacting   to clarify several issues regarding
temporal coding.

http://neuronline.sfn.org/SFN/SFN/Home/Default.aspx (require membership)
www.linkedin.com/groups/Computational-Neuroscience-1376707 (require
 Dorian Aur
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://www.neuroinf.org/pipermail/comp-neuro/attachments/20110811/22155b8d/attachment-0001.html

More information about the Comp-neuro mailing list