README.txt

*****************************************
Grid cell spatial pattern models
  eric zilli 
*****************************************

Version 1.007.

This package contains MATLAB implementations of grid cell models drawn
from the literature in the years 2005 to 2011.

This package does not yet have *all* models that can produce the
grid pattern. Right now these are missing and here's why:
* Welday et al. 2011 just came out. It is essentially the Blair et al. 2008
    model, though they used a cool approach of writing the activity in terms
    of the interference envelope, but BlairEtAl2008.m suffices to test it
    for now. More importantly, they were great enough to share their code
    on ModelDB http://senselab.med.yale.edu/modeldb/ShowModel.asp?model=129067
    Someone buy them a beer.
* O'Keefe and Burgess 2005 and McNaughton et al. 2006 talked about
    models but didn't give enough details as to describe a unique model to
    implement.
* XXXX in print and XXXX in review. I'm aware of 2 developmental models
    but cannot describe articles not yet published. I'm intending to name
    the paper in a manner to suggest it covers only models published
    through 2011 so if these are published past December technically I don't
    have to implement them! I probably will, though. I'm a loner, Dottie.
    A rebel.
* Zhang 1996, Samsonovich and McNaughton 1997, Conklin and Eliasmith 2005
    to name three, are older place cell models, but two have toroidal
    topologies (and Zhang 1996 mentioned it) and so produce rectangular
    grids. Technically observed entorhinal grids are almost always
    hexagonal, but O'Keefe and Burgess 2005 was a rectangular grid
    model, for instance, and I still mention that one. At the very least
    Zhang 1996's ring model is worth implementing for its comprehensive
    theoretical approach and Conklin and Eliasmith 2005 because I've
    long wanted to better understand their framework for deriving
    attractor networks analytically.
* Amari 1977 considered a neural field (a continuous attractor network type
    system) in which multi-peaked solutions could occur. This was not quite a
    grid cell model, but a link was pointed out by Spencer et al. "The Dynamic
    Field Theory and Embodied Cognitive Dynamics"   
    (http://homepage.psy.utexas.edu/homepage/group/loveLAB/love/classes/models/SC.pdf),
    footnote 3. Verily I quoth,
      "3. One of the attractor states Amari identified was a [...] spatially
      periodic state, where the period varies with the average stimulation to
      the field. We suspect this state is involved in extracting the axes of
      symmetry we have probed in our work on the development of spatial recall
      (see, e.g., Schutte, Spencer, & Schöner, 2003). This state might also have
      ties to the pattern of “grid” cells observed in entorhinal and perirhinal
      cortex (Hafting, Fyhn, Molden, Moser, & Moser, 2005)."


*****************************************
* Version history
*****************************************

The newest version of this package can be found at
http://people.bu.edu/zilli/gridmodels.html or by emailing
me should that page disappear. I'm not a huge fan of ModelDB,
but I might throw a backup up there too just in case.

v1.007 20120127
  Removed a comment in Guanella et al. 2007 that incorrectly said
    normalization was not needed in the model.

v1.006 20120113
  Cleaning up README.txt a little.

v1.005 20111230
  Cleaning up README.txt a little.

v1.004. 20111214
  Added model FuhsTouretzky2006_development.m

v1.003. 20111214
  Downsampled data in ZilliHasselmo2010_voltage_traces.mat
    used for FigureMaintain.m for smaller package size.
  Fixed a bug in BurgessEtAl2007.m, Burgess2008_bat.m,
    GiocomoEtAl2007.m, Hasselmo2008_bat.m, NavratilovaEtAl2011.m,
    and ZilliHasselmo2010.m where spikes were being drawn slightly
    offset from the trajectory.
  Cleaned out some old comments in FigureMaintain.m

v1.002. 20111213
  Added model KropffTreves2008.m

v1.001. 20111211.
  Added model GiocomoEtAl2007.m
  Added model ZilliHasselmo2010.m
  Added data file simple_model_RS2_FI_Jan09_n1.mat
  Added data file simple_model_RS2sn_FI_Jan09_n250.mat

v1.0. 20111208.
  First public release.


*****************************************
* Errata
*****************************************

None yet.


*****************************************
* Scripts that generate Figures
*****************************************

These figures may be useful when talking about the models.

The figure scripts are pretty messy. The figures are all drawn
directly in MATLAB and that involves drawing a lot of rectangles
and lines, plotting data, drawing text, and manipulating axis
properties. Use MATLAB's cell titles feature if you have a recent
version to more easily navigate these files (the little button in
the editor with two percentage signs and a down arrow).

The figures generated by these scripts were saved as a .pdf
using the great script export_fig from MATLAB Central which gives
very pretty output files. The figures use Arial, but MATLAB
cannot produce output with Arial, so Illustrator 
can be used to convert the fonts (use Type->Find Font... to quickly
replace the fonts). The resulting output is slightly uglier than
the original pdf.

*** Figure_AttractorWeights.m
 This generates Figure 3, which demonstrates the functioning of three CAN
models. The data plotted was generated using the scripts
FuhsTouretzky2006.m, GaussierEtAl2007.m, and BurakFiete2009.m. Commented
out code at the beginning of these scripts will calculate and save the
files necessary to plot Figure 3.
(This data is included pre-generated in the files Fu06_WeightFigure_vars.mat,
Gu07_WeightFigure_vars.mat, and Bu09_WeightFigure_vars.mat)

*** Figure_BatGrid.m
 This generates a figure not included in the manuscript. The figure was
designed to demonstrate how different interference models are in fact
consistent with the bat grid cell data reported by Yartsev et al. 2011.
This script calls _bat.m variations on three models, BurgessEtAl2007_bat.m,
Burgess2008_bat.m, and Hasselmo2008_bat.m. Generally these scripts
change two things about the model: the baseline frequency is set to 0 Hz
and oscillators are modified so that their frequency only increases in
response to velocity input.

*** Figure_Grid.m
 This generates Figure 1, which demonstrates various conceptualizations
of the hexagonal grid pattern.

*** Figure_Maintain.m
 This generates Figure 2, which demonstrates various ways that the different
models encode linear and planar positions.
(Uses files generalGridPattern.mat, Fu06_WeightFigure_vars.mat,
Gu07_WeightFigure_vars.mat, and Bu09_WeightFigure_vars.mat)

*** Figure_Readout.m
 This generates Figure 4, which demonstrates the read-out rules that have
been used in the temporal interference models.
 
*** drawSimpleRing.m
 This function is used to draw simple ring attractors to the current axes.
It accepts parameters specifying how many nested rings of cells to draw,
how many cells (drawn as circles) in each ring, how large the cells should
be drawn, which cell should be drawn as active, what colors to draw each
cell, and whether or not to draw an arrow around the ring indicating
the motion of a biased ring attractor. It's a bit messy and poorly written,
but might be useful to others (at least as a starting point).


*****************************************
* Model scripts
*****************************************

The names are pretty self-explanatory so I'll only describe the variations.
More details are available in the original manuscripts (see references
below), the scripts themselves, and my manuscript.

*** BlairEtAl2007.m

*** BlairEtAl2008.m

*** BlairEtAl2008_2D.m
  This is a 2D version of their 1D model.

*** BurakFiete2009.m

*** Burgess2008.m

*** Burgess2008_bat.m
  This is a version of the model configured to be consistent
  with the Yartsev et al. 2011 bat data. It accepts command
  line options so that FigureBatGrid.m can pull out the
  data it needs.

*** BurgessEtAl2007.m

*** BurgessEtAl2007_bat.m
  This is a version of the model configured to be consistent
  with the Yartsev et al. 2011 bat data. It accepts command
  line options so that FigureBatGrid.m can pull out the
  data it needs.

*** BurgessEtAl2007_precession.m
  This implements my guesses as to what Burgess meant in his
  precession discussion

*** FuhsTouretzky2006.m

*** FuhsTouretzky2006_development.m
  This implements Fuhs and Touretzky (2006)'s developmental method
  for learning the symmetric component of their weight matrix using
  sinusoidal gratings of activity analogous to retinal waves.

*** GaussierEtAl2007.m

*** GuanellaEtAl2007.m

*** GuanellaEtAl2007_no_twist.m
  This demonstrates how a single-bump torus can produce hexagonal
  fields by skewing the input velocities.

*** Hasselmo2008.m

*** Hasselmo2008_bat.m
  This is a version of the model configured to be consistent
  with the Yartsev et al. 2011 bat data. It accepts command
  line options so that FigureBatGrid.m can pull out the
  data it needs.

*** HasselmoBrandon2008.m

*** MhatreEtAl2010.m

*** NavratilovaEtAl2011.m

*** ZilliHasselmo2010.m


*****************************************
* Data files
*****************************************

To plot some of the figures, data generated in simulations is required.
The following data files are included.

*** 11207-21060501+02_t6c1.mat
 This contains rat trajectory from Sargolini et al. 2006. It is used
as the input to the Mhatre et al. 2010 model.

*** BlairEtAl2007_Readout.mat
 This contains a variable of their moire interference mechanism as seen
in my Figure 1.

*** Bu09_WeightFigure_vars.mat
 This contains network activity and synaptic input variables used in
Figures 2 and 3.

*** Fu06_WeightFigure_vars.mat
 This contains network activity and synaptic input variables used in
Figures 2 and 3.

*** generalGridPattern.mat
 This contains a variable that is an image of hexagonally arrayed
fields to show how the idealized hexagonal grid pattern looks. It
was generated by summing three 2D cosine gratings a la Blair et al.
(2007). Used in Figures 1 and 2.

*** Gu07_WeightFigure_vars.mat
 This contains network activity and synaptic input variables used in
Figures 2 and 3.

*** HaftingTraj_centimeters_seconds.mat
 This contains trajectory from Hafting et al. 2005 of a rat running
around an open field for about 591 seconds (a tad under 10 minutes).

*** simple_model_RS2_FI_Jan09_n1.mat
 This contains an FI curve (generated by SI_simple_model_FI_relation.m
 in the Zilli and Hasselmo scripts on ModelDB) of a type 2 excitable
 regular spiking simple model neuron for use in ZilliHasselmo2010.m.

*** simple_model_RS2sn_FI_Jan09_n250.mat
 This contains an FI curve (generated by SI_simple_model_FI_relation.m
 in the Zilli and Hasselmo scripts on ModelDB) of a network of 250 
 all-to-all connected type 2 excitable regular spiking simple model
 neurons connected with delta synapses and injected with a realistic
 level of noise for use in ZilliHasselmo2010.m.

*** Spatial_interference.mat
 This contains an image of two sets of spatial bands at a 60 degree
angle overlapping to produce a hexagonal pattern. Used in Figure 1.

*** ZilliHasselmo2010_voltage_traces.mat
 This contains voltage traces of VCOs from the Zilli and Hasselmo
model to show phase-synchronized spiking. Used in Figure 2.


*****************************************
* Model speeds
*****************************************

The simulations of the models vary greatly in speed, i.e. how long you have
to sit there waiting for something interesting to happen. The simulations
were not thoroughly optimized for speed, so a bit of improvement is
possible in some models, but overall these times do fairly represent the
complexity of the various models. (Note that, e.g. MhatreEtAl2010.m runs
fairly quickly on a per-run basis, but it is a developmental so many runs
are required, making it slower than other models where much shorter
simulations suffice to show that the models are working)

Simulation speeds (with graphics disabled, i.e. livePlot=0):
* Instant:
  *** BlairEtAl2007.m - All we have to do is make two theta grids and add
          them together.
  *** BlairEtAl2008.m - I wrote this one fully vectorized so it is very
          fast. Hasselmo2008.m, probably GaussierEtAl2007.m, and the
          Burgess models could also be easily vectorized in this manner
          for faster simulations.

* Fast:
  *** HasselmoBrandon2008.m - (about 1 s to run 200 s of simulation)
  *** Burgess2008.m - (about 2 s for a 200 s run)
  *** BurgessEtAl2007.m - (about 2 s for a 200 s run)
  *** Hasselmo2008.m - (about 6 s for a 200 s run, faster if
          spikeTimes/spikeCoords/spikePhases were pre-alloc'd)

* Medium:
  *** GaussierEtAl2007.m - (about 20 s for a 200 s run)
  *** GuanellaEtAl2007.m - (about 60 s for a 200 s run)

* Slow:
  *** ZilliHasselmo2010.m - (1 cell per oscillator; about 55 s for a 20 s run)
  *** ZilliHasselmo2010.m - (250 cells per oscillator; about 100 s for a 20 s run)
  *** FuhsTouretzky2006.m - (about 220 s for a 20 s run)
  *** NavratilovaEtAl2011.m - (about 766 s for a 20 s run)

* Ent-like:
  *** BurakFiete2009.m - (about 170 s for a 1 s run, not counting the
          hour or so to generate W)

* Developmental (not fair to compare to the other models):
  *** MhatreEtAl2010.m - (about 360 s for a 1200 s run but 5-20 are needed)
  *** KropffTreves2008.m - (about 3560 s for a 500,000 step run)


*****************************************
* Irreleventia
*****************************************

To make this collection, I had to make a number of decisions about what
constitutes a grid cell model and what constitutes implementing it.

Modeling papers generally show how some mechanism can produce some
apparently unrelated phenomenon, and presumably the authors also assume
that any obvious variation on the mechanism is also described by their
model, not uncommonly making explicit the variations they considered most
interesting. This provides the first issue: if a paper describes many
variations on a model, should only some or all be implemented? My bias was
toward implementing the simpler versions, or more than one if many were
easy to do.

Many models contain not just one mechanism, but multiple interacting
mechanisms that produce the phenomenon of interest. Not uncommonly,
modelers do not model all the mechanisms, but rather assume some
mechanisms work perfectly and calculate what they would be expected to do
rather than specifying how it might be done. This simplifies research
by saving programmers from redoing work that has already been done, but
can be dangerous in that it is easy to assume impossible things. I have
tried to follow the lead of the authors, generally implementing what they
implemented and not implementing what they didn't. Be warned that the
ability of any model to produce correct-looking output does not mean it
is doing it in a sane or robust manner (even if it is published!).

The points above apply to all modeling, but this review was focused on grid
cells so I had to identify the grid cell models to review. I hope I found
and discussed all of the substantive models directly addressing grid cells,
but most grid cell models derive from earlier place cell models that
produced a repeating grid of fields. Thus many place cell models were
or contained grid cell models, even though grid cells were not known at the
time (or were just being discovered, e.g. Conklin and Eliasmith 2005).

*****************************************
* Model references
*****************************************

These are the papers I know of containing grid cell models or place cell
models that could produce grid cells. I'm happy to share pdfs of any or
all if you can't access them.

Blair, H. T., Gupta, K., & Zhang, K. (2008). Conversion of a phase- to
rate-coded position signal by a three-stage model of theta cells, grid
cells, and place cells. Hippocampus, 18, 1239-1255.

Blair, H. T., Welday, A. W., & Zhang, K. (2007). Scale-invariant memory
representations emerge from Moire interference between grid fields that
produce theta oscillations: A computational model. J Neurosci, 27,
3211-3229.

Burak, Y., & Fiete, I. R. (2009). Accurate path integration in continuous
attractor network models of grid cells. PLoS Computational Biology, 5(2).

Burgess, N. (2008). Grid cells and theta as oscillatory interference:
theory and predictions. Hippocampus, 18(12), 1157-74.

Burgess, N., Barry, C., & O’Keefe, J. (2007). An oscillatory interference
model of grid cell firing. Hippocampus, 17, 801-812.

Conklin, J., & Eliasmith, C. (2005). A controlled attractor network model
of path integration in the rat. J Comput Neurosci, 18(2), 183-203.

Fuhs, M. C., & Touretzky, D. S. (2006). A spin glass model of path
integration in rat medial entorhinal cortex. J Neurosci, 26, 4266-4276.

Gaussier, P., Banquet, J. P., Sargolini, F., Giovannangeli, C., Save, E.,
& Poucet, B. (2007). A model of grid cells involving extra hippocampal 
path integration, and the hippocampal loop. J Integrated Neurosci, 6(3),
447-476.

Giocomo, L. M., Zilli, E. A., Frans´en, E., & Hasselmo, M. E. (2007).
Temporal frequency of subthreshold oscillations scales with entorhinal
grid cell field spacing. Science, 23, 1719-1722.

Guanella, A., Kiper, D., & Verschure, P. (2007). A model of grid cells
based on a twisted torus topology. Int J Neural Syst, 17, 231-240.

Hasselmo, M. E. (2008). Grid cell mechanisms and function: Contributions
of entorhinal persistent spiking and phase resetting. Hippocampus, 18,
1213-1229.

Hasselmo, M. E., & Brandon, M. P. (2008). Linking cellular mechanisms to
behavior: Entorhinal persistent spiking and membrane potential
oscillations may underlie path integration, grid cell firing and episodic
memory. Neural Plasticity, 658323.

Kropff, E., & Treves, A. (2008). The emergence of grid cells: Intelligent
design or just adaptation. Hippocampus, 18(12), 1256-1269.

McNaughton, B. L., Battaglia, F. P., Jensen, O., Moser, E. I., & Moser,
M. B. (2006). Path integration and the neural basis of the ‘cognitive
map’. Nat Rev Neurosci, 7, 663-678.

Mhatre, H., Gorchetchnikov, A., & Grossberg, S. (2010). Grid cell
hexagonal patterns formed by fast self-organized learning within
entorhinal cortex. Hippocampus, 1098-1063. Available from
http://dx.doi.org/10.1002/hipo.20901

Navratilova, Z., Giocomo, L. M., Fellous, J.-M., Hasselmo, M. E., &
McNaughton, B. L. (2011). Phase precession and variable spatial scaling 
in a periodic attractor map model of medial entorhinal grid cells with 
realistic after-spike dynamics. Hippocampus, 1098-1063. Available from 
http://dx.doi.org/10.1002/hipo.20939

O’Keefe, J., & Burgess, N. (2005). Dual phase and rate coding in 
hippocampal place cells: Theoretical significance and relationship to 
entorhinal grid cells. Hippocampus, 15, 853-866.

Samsonovich, A., & McNaughton, B. L. (1997). Path integration and 
cognitive mapping in a continuous attractor neural network model. Journal 
of Neuroscience, 17(15), 5900-5920.

Welday, A. C., Shlifer, I. G., Bloom, M. L., Zhang, K., & Blair, H. T. 
(2011). Cosine directional tuning of theta cell burst frequencies: 
Evidence for spatial coding by oscillatory interference. Journal of 
Neuroscience, 31(45), 16157-16176. Available from 
http://www.jneurosci.org/content/31/45/16157.abstract

Zhang, K. (1996). Representation of spatial orientation by the intrinsic 
dynamics of the head-direction cell ensemble: a theory. Journal of 
Neuroscience, 16(6), 2112-2126.

Zilli, E. A., & Hasselmo, M. E. (2010). Coupled noisy spiking neurons as 
velocity-controlled oscillators in a model of grid cell spatial firing. 
Journal of Neuroscience, 30, 13850-13860.


*****************************************
* Epilogue
*****************************************

I started as a grad student in September 2003 in Michael Hasselmo's lab,
which was at the time largely focused on theta rhythm in the hippocampus
(beginning to move into prefrontal cortex). I therefore started in
neuroscience by catching up on the experimental and theoretical work done
on the hippocampus, putting me in a good position to be surprised when
suddenly grid cells appeared in 2005. That may color my opinions in
general, as I learned that just when everyone thinks one thing is figured
out, a completely unexpected aspect of the problem is discovered, so it
is hard to be too certain about anything!

(Of course, had I known the literature even better, I'd have known repeating
fields were produced by many earlier models of place cells and perhaps
would not have found them as surprising!)

I remember first seeing Hafting's 2005 paper in a lab meeting and staying
after and staring at the hexagonal pattern, trying to figure out how a cell
could know when the animal had reached positions equally spaced out in a
hexagonal grid. Afterward I'd spend time thinking about grid cells, as I
think many people in the field did. Many aspects were quickly clear: the
repeating code at multiple scales was an efficient way of representing
positions (or times, as I examined in unpublished work in my PhD thesis,
though I have since discovered a number of earlier models that did the same),
but there was no easy way to take a vector of grid activities at a starting
point and an ending point and calculating the distance or angle, due to the
repeating nature of the code. They provided an efficient way to mentally
move through a representation of space (I used grid cells with reinforcement
learning to solve a task that required mentally navigating space in
unpublished work following Zilli and Hasselmo 2008, Hippocampus), but that
doesn't much suggest a mechanism. I did find, though, as many did, that
Mexican-hat connectivity on an attractor sheet will organize bumps into a
hexagonal grid. Also, if you screw with the parameters of that network you
can get some really psychedelic patterns to dance across the cells.

Around 2006 the renowned Lisa Giocomo was recording from stellate cells in
entorhinal cortex layer II and found a gradient of resonant frequencies
along the dorsoventral axis of ECII (testing a prediction from O'Keefe and
Burgess 2005). While she was collecting that data, I tried to
get a model to produce grid cells and take her data into account. Like
Kropff and Treves (2008), I made a learning rule (mine resonance-based)
that allowed a set of grid-cells--to--be to organize weights from
place cells at hexagonally-arrayed positions so that low resonance frequencies
would produce larger spacings, as in the data. This model was never
published because I could never get the thing to actually path integrate
once these weights were learned. But John White did call my mechanism
clever, though I suppose that could have been a euphemism!

Before long Burgess and friends had a poster on an oscillatory interference
grid cell model, and he sent Mike Hasselmo a jpg of their poster, and he
passed it on to me. I implemented that model as described on the poster
(which was written in constant-velocity form, so Mike and I subsequently
had to figure out a good way to write it with a variable velocity).

When Lisa wrote up her ECII resonance data, we included the oscillatory
interference model in the paper (I'd already given up on my self-organizing
model) and related her data to that.

I'd finished my PhD thesis at that point (Sept 07) and spent the rest
of that year and 2008 finishing up some loose ends from my thesis that were
published as Zilli and Hasselmo (2008a Hippocampus, 2008b PLoS ONE, 2008c
Frontiers Comput Neurosci), all regarding reinforcement-learning--related
gridworld tasks with various memory systems.

I returned from gridworld to grid cell world in 2009 by examining a
common problem identified with interference models: the actual oscillations
in the brain are considerably more noisy than the noiseless ones used
in grid cell models up to that time. We looked at this quantitatively
by analyzing recordings of various neural oscillators and calculated how
long a grid pattern would remain stable if those oscillations were used.
Most of the cells would have had stability times on the order of fractions
of a second, with the best times at perhaps 1.5 s.

These results did not particularly support interference models so, I was
encouraged not to actually give those low stability times explicitly in the
abstract because it would look bad for the models. Luckily the message got
across anyhow, but I regret changing that.

Of course, that data did not doom the interference models, because there
are many other oscillating fish in the sea, and, in particular, Mike
thought interactions among oscillators could reduce the effects of noise.
In Zilli and Hasselmo (2010) we showed that indeed this was true: by
coupling noisy oscillators, the network as a whole produced more regular
oscillations than the cells that comprised it. It did appear, however,
that quite a large number of neurons must be coupled together to get
the desired high stability times. I expect our results are qualitatively
true of real neurons, but the quantitative question is harder to answer.
If I needed 1,000 model neurons in a network to be stable for some
desired amount of time, it is not clear whether 1,000 real neurons would
also suffice, or perhaps whether it would be 100 or 10,000. The situation
could thus be better or worse than it appears.

But it is essentially impossible that any one current model
is exactly true, and simultaneously most features of the models are too
generally true of all circuits in the brain for them to be disproven
(like the way pattern separation and pattern completion are general
properties of neural networks, yet are specifically attributed to the HC).
This also has implications for testing the predictions of the models. It
is an excellent use of the models to guide research by determining
conflicting predictions of distinct models, but all of the models must
be understood in the first place to identify the predictions that are
in fact in conflict.

For example, some may claim that both the resonance gradient of Giocomo
et al. (2007) and more recent results by ex-local, now SoCal heartthrob
Mark Brandon et al. (2011; showing inactivation of the medial septum
disrupts grid cell firing) support the interference models. However, the
two results are not mutually consistent. Giocomo et al. (2007)'s model was
specifically modified to not use theta as the baseline oscillation, whereas
Brandon et al. (2011)'s data was suggested as supporting interference
models because those models use theta as a baseline. In both cases
interesting new results were produced by testing predictions of models,
but those data do not uniquely support any single model or class of
models.

That brings us to the present, three weeks before Christmas Eve and
four weeks until 2012: the future! It is worth considering future
directions for grid cell research.

Common to all the path integration models is the calculation of
directional velocity inputs. These are worth tracking down. I'd wager
a good place to look is the cerebellum, since there is a very direct
path from the otoliths in the inner ear that sense certain kinds of
motion to the cerebellum (first to the nodus/uvula then to the
fastigial nucleus) and then right to the medial temporal lobe, and,
in all likelihood, entorhinal cortex. That cerebellar nucleus seems in
prime position to provide the body velocity signal needed for the
models, though I wouldn't be surprised if it were found in
the thalamus too/instead (see Welday et al. 2011).

Another thing worth examining is whether grid cells always reflect
the animal's current position. It is reasonable to assume that when
mentally navigating an environment, the grid cell network could
be constantly changing to reflect the mental navigation, rather than
the animal's current position. Similarly, grid cell patterns might
show the "replay" phenomena reported in place cells (though I don't
think people do the statistics appropriately on those so I'm not quite
convinced replay is actually a real thing, even though it does make
complete sense).

Another question that has not been thoroughly address is, apart from
edge effects that clearly distort the grid, how globally consistent
is the hexagonal pattern? Using the spatial power spectral density
rather than the spatial autocorrelation alone may be a useful tool
to measure this. Brun et al. (2008) showed rather inconsistent
spacing on the long linear track, including directional-changes in
field spacing. Alternatives to perfectly consistent spacing include
grid fields that are roughly hexagonal but consistently randomized
in spacing between neighboring fields, roughly hexagonal but in
non-parallel lines, or "fractured" grids where fields in one half
of the environment have consistent spacing and similarly in the
other half, but separated by a "fault line" where the two patterns
do not agree. These types of patterns can be observed in
developmental models like Mhatre et al. (2010) or in spatial
interference models using imperfect patterns like Turing stripes
(McNaughton et al. 2006).

Similarly, how globally consistent is the crosscorrelation of two
cells? The independent positional code of linear coding models could
easily produce inconsistent crosscorrelations, whereas planar coding
models would predict more consistent patterns of nearby cells.
Sensory associations that reset grid cells would be expected to
increase the consistency, so an observation of inconsistency could
be useful information for understanding how the pattern is generated.
However, 2D continuous attractor models can transiently sustain flawed 
patterns, so the crosscorrelation may be best performed on shorter
time windows.

Other question:
* Do grid cells fire on every pass through a field? Hafting et al.
  2005's Figure 6a suggests they may not, at least early in exposure to
  a novel environment: In the 0-1 minute plot the animal twice passes
  through two fields on the left edge but the cell does not fire. Is that
  merely an edge effect?
* Why do environmental edges distort the grid? And if the regularity
  of the grid is key to its use as a spatial code, is any spatial
  ability impaired at edges where this distortion occurs?
* When a grid cell has a head direction preference, is it always
  aligned with one of the three directions along which the grid
  fields line up?
* Do pentagonal or heptagonal field groupings actually occur or
  are they artifacts from edge distortions and/or the animal
  repeatedly taking two distinct trajectories through a field
  (and so seeming to split one field into two)?
* Is the amount of phase precession observed within a field correlated
  with the ratio of field width to field spacing? Temporal interference
  models predict that firing phase range = 360*(field width)/(field spacing).
  This phase range applies only to the range where the spikes are
  truly precessing, not the second component (Yamaguchi et al. 2001)
  as the animal exits the field where the cell fires across most phases.
* Do dorsal grid cells develop before ventral grid cells?
  Experience-dependent learning models might be able to more quickly
  organize grid cells with smaller field spacing.

If any further information is needed, write to eric dot zilli at
gmail dot you had better not spam me! dot com.
 
eric zilli - 20111204