The GCAL model (gain control, adaptation and lateral) is a general purpose cortical developmental model of topographic map formation in the primary visual cortex (V1), driven by the presentation of visual patterns on a retinal sheet. This model accounts for receptive field formation, the emergence and of orientation selectivity across the cortical surface as well as contrast-invariant orientation tuning. The GCAL model is robust to large changes in the contrast of the training patterns that may flucuate between widely varying statistical distributions. The resulting orientation maps are robust and stable yet adaptive development that reflects the statistics of the input.
The GCAL model and related submodels (L,AL and GCL) are fully described by Stevens et al. (2013):
@article{Stevens2013, author = {Stevens, Jean-Luc R. and Law, Judith S. and Antol\'{i}k, J\'{a}n and Bednar, James A.}, title = {Mechanisms for Stable, Robust, and Adaptive Development of Orientation Maps in the Primary Visual Cortex}, journal = {The Journal of Neuroscience}, volume = {33}, number = {40}, pages = {15747-15766}, year = {2013}, doi = {10.1523/JNEUROSCI.1037-13.2013}, url = {http://www.jneurosci.org/content/33/40/15747.full} }
If you can double-click on this text and edit it, you are in a live IPython Notebook environment where you can run the code and explore the model. Otherwise, you are viewing a static (e.g. HTML) copy of the notebook, which allows you to see the precomputed results only. To switch to the live notebook, see the notebook installation instructions.
This IPython notebook constructs the definition of the GCAL model and related submodels (L,AL and GCL), for use with the Topographica simulator. The resulting model file is used by the accompanying notebook to launch all the necessary simulations and collate the necessary data to plot all the figures published in Stevens et al. (2013).
Running this notebook defines the model file source (gcal.ty) while the model is loaded into memory for inspection, allowing you to see how the model components are specified and put together. Examples of the retinal training patterns and model weigths are shownn as well as the orientation map after running 3000 iterations.
A static version of this notebook may be viewed online. To run the live notebook and explore the model interactively, you will need both IPython notebook and Topographica. The installation instructions and software dependencies are listed here
Note that the gcal.ty model file exported by the notebook may be run without IPython using Topographica as follows:
./topographica -g ./gcal.ty
Once you have exported the gcal.ty
file, you can also conveniently load the model into Topographica and launch the GUI from here. Note that you will need topo
on the path - the following code cell imports the modules needed in this notebook and adds topo
to the path, assuming you are running this notebook in the stevens.jn13
directory:
import os, sys, math
sys.path += [os.path.abspath(os.path.join('..','..'))]
import topo
import imagen
import lancet
import pandas
from jn13_figures.lib import misc
prompt_msg = 'Would you like to explore the GCAL model with the Topographica GUI? (y/n): '
if (raw_input(prompt_msg) if True else 'n') == 'y': # Set False to disable prompt.
misc.open_gui()
if os.path.isfile('gcal.ty'):
execfile('gcal.ty')
else:
print "GCAL not loaded: 'gcal.ty' file not found!"
This section includes some utilities used later in this notebook, the model docstring and imports and the model parameters. Anything marked __GCAL_TY__ will go to the .ty file later,
__GCAL_TY__ = True # Code cells marked at the top with this identifier will be exported to gcal.ty
__GCAL_TY__
# -*- mode: python;-*-
"""
GCAL
Simple but robust single-population V1 model orientation map from:
Jean-Luc R. Stevens, Judith S. Law, Jan Antolik, and James A. Bednar.
Mechanisms for stable, robust, and adaptive development of orientation
maps in the primary visual cortex.
@article{Stevens02102013,
author = {Stevens, Jean-Luc R. and Law, Judith S. and Antolik, Jan and Bednar, James A.},
title = {Mechanisms for Stable, Robust, and Adaptive Development of Orientation Maps in the Primary Visual Cortex},
journal = {The Journal of Neuroscience}
volume = {33},
number = {40},
pages = {15747-15766},
year = {2013},
doi = {10.1523/JNEUROSCI.1037-13.2013},
url = {http://www.jneurosci.org/content/33/40/15747.full}
}
"""
from math import pi
import os
import numpy
import param
from topo import learningfn,numbergen,transferfn,pattern,projection,responsefn,sheet
import topo.learningfn.optimized
import topo.learningfn.projfn
import topo.transferfn.optimized
import topo.pattern.random
import topo.pattern.image
import topo.responsefn.optimized
import topo.sheet.lissom
import topo.sheet.optimized
import topo.transferfn.misc
from topo.base.arrayutil import DivideWithConstant
from topo.transferfn.misc import HomeostaticResponse
from topo.misc.commandline import global_params as p
These are the parameters that define the quality of the simulation, the exact model loaded (L,AL,GCL or GCAL), the dataset used and other key model and training parameters:
__GCAL_TY__
p.add(
input_seed = param.Integer(default=42, doc="""
The random seed to set the input patterns"""),
dataset=param.ObjectSelector(default='Gaussian',objects=
['Gaussian','Nature', 'VGR'],doc="""
Set of input patterns to use::
:'Gaussian': Two-dimensional Gaussians
:'Nature': Shouval's 1999 monochrome 256x256 images.
:'VGR': Simulated vertical goggle rearing
(anisotropically blurred Shouval)"""),
num_inputs=param.Integer(default=2,bounds=(1,None),doc="""
How many input patterns to present per unit area at each
iteration, when using discrete patterns (e.g. Gaussians)."""),
aff_strength=param.Number(default=1.5,bounds=(0.0,None),doc="""
Overall strength of the afferent projection to V1."""),
exc_strength=param.Number(default=1.7,bounds=(0.0,None),doc="""
Overall strength of the lateral excitatory projection to V1."""),
inh_strength=param.Number(default=1.4,bounds=(0.0,None),doc="""
Overall strength of the lateral inhibitory projection to V1."""),
aff_lr=param.Number(default=0.1,bounds=(0.0,None),doc="""
Learning rate for the afferent projection to V1."""),
exc_lr=param.Number(default=0.0,bounds=(0.0,None),doc="""
Learning rate for the lateral excitatory projection to V1."""),
inh_lr=param.Number(default=0.3,bounds=(0.0,None),doc="""
Learning rate for the lateral inhibitory projection to V1."""),
area=param.Number(default=1.0,bounds=(0,None),
inclusive_bounds=(False,True),doc="""
Linear size of cortical area to simulate.
2.0 gives a 2.0x2.0 Sheet area in V1."""),
retina_density=param.Number(default=24.0,bounds=(0,None),
inclusive_bounds=(False,True),doc="""
The nominal_density to use for the retina."""),
lgn_density=param.Number(default=24.0,bounds=(0,None),
inclusive_bounds=(False,True),doc="""
The nominal_density to use for the LGN."""),
cortex_density=param.Number(default=49.0,bounds=(0,None),
inclusive_bounds=(False,True),doc="""
The nominal_density to use for V1."""),
# Noisy disk parameters
retinal_waves=param.Integer(default=0,bounds=(0,None),doc="""
How many retinal wave (noisy disk) presentations before the
chosen dataset is displayed. Zero for no noisy disk
presentations otherwise 6000 has been typically used.."""),
percent_noise = param.Number(default=50, doc="""
The percentage of the noise strength to the disk strength for noisy disks"""),
contrast=param.Number(default=20, bounds=(0,100),
inclusive_bounds=(True,True),doc="""
Brightness of the input patterns as a contrast (percent)."""),
# Toggle between the L, GCL, AL and GCAL models
gain_control = param.Boolean(default=True, doc="""
Whether or not gain-control (divisive lateral inhibitory) is to be
applied in the LGN"""),
homeostasis = param.Boolean(default=True, doc="""
Whether or not the homeostatic adaption should be applied in V1"""),
t_init = param.Number(default=0.20, doc="""
The initial V1 threshold value. This value is static in the L and GCL models
and adaptive in the AL and GCAL models.""")
)
On every training iteration, a stimulus is presented on the retina. To demonstrate the robustness of the GCAL model, we used four types of image:
__GCAL_TY__
### Input patterns
# Scale of 1.0 is equivalent to 100% contrast.
contrast_scale = p.contrast / 100.0
if p.dataset=="Gaussian":
total_num_inputs=int(p.num_inputs*p.area**2)
inputs=[pattern.Gaussian(x=numbergen.UniformRandom(lbound=-(p.area/2.0+0.25),
ubound= (p.area/2.0+0.25),seed=p.input_seed+12+i),
y=numbergen.UniformRandom(lbound=-(p.area/2.0+0.25),
ubound= (p.area/2.0+0.25),seed=p.input_seed+35+i),
orientation=numbergen.UniformRandom(lbound=-pi,ubound=pi,seed=p.input_seed+21+i),
size=0.088388, aspect_ratio=4.66667, scale=contrast_scale)
for i in xrange(total_num_inputs)]
combined_inputs = pattern.SeparatedComposite(min_separation=0,generators=inputs)
misc.view_patterns(combined_inputs,radius=p.area/2.0+0.25+0.375+0.5)
images_dir = param.resolve_path('images', path_to_file=False)
images = lancet.FilePattern('Nature', os.path.join(images_dir, 'shouval','combined*.png'))
lancet.ViewFrame(images.dframe.head(3).T)
isotropic_kernel = imagen.Gaussian(aspect_ratio=1.0, xdensity=128, ydensity=128,
size=0.05, orientation=math.pi/2.0)
misc.generate_GR(images_dir, isotropic_kernel, name='blur')
images = lancet.FilePattern('blur', os.path.join(images_dir, 'blur','combined*_blur.png'))
convolve_dframe = pandas.DataFrame({'blur':[isotropic_kernel]}).append(images.dframe.head(2))
lancet.ViewFrame(convolve_dframe.T)
misc.generate_GR(images_dir, name='VGR')
images = lancet.FilePattern('VGR', os.path.join(images_dir, 'VGR','combined*_VGR.png'))
anisotropic_dframe = pandas.DataFrame({'VGR':[misc.anisotropic_kernel]}).append(images.dframe.head(2))
lancet.ViewFrame(anisotropic_dframe.T)
__GCAL_TY__
if p.dataset in ["Nature", "VGR"]:
# Do not randomly rotate GR patches (otherwise horizontal bias is meaningless)
patch_orientation = 0.0 if p.dataset=="VGR" else numbergen.UniformRandom(lbound=-pi,ubound=pi,seed=p.input_seed+65)
if p.dataset=="Nature":
image_filenames= [param.resolve_path("images/shouval/combined%02d.png"%(i+1)) for i in xrange(25)]
else:
# Actual ordering used in the paper for historical reasons,
# makes little difference if ordering=range(1,26) is used instead.
ordering = [1,25,8,24,22,23,4,15,17,13,5,20,18,14,3,6,12,10,21,7,9,2,16,11,19]
image_filenames = [param.resolve_path("images/VGR/combined%02d_VGR.png" % i) for i in ordering]
inputs=[pattern.image.FileImage(filename=f,
scale=contrast_scale,
size=10.0,
x=numbergen.UniformRandom(lbound=-0.75,ubound=0.75,seed=p.input_seed+12),
y=numbergen.UniformRandom(lbound=-0.75,ubound=0.75,seed=p.input_seed+36),
orientation=patch_orientation) for f in image_filenames]
combined_inputs = pattern.Selector(generators=inputs)
__GCAL_TY__
noise_ratio = (p.percent_noise / 100.0)
disk_scale= 1.0 / (1.0 + noise_ratio)
rand_scale= noise_ratio / (1.0 + noise_ratio)
disks_inputs=[topo.pattern.Composite(operator=numpy.add,
scale=contrast_scale,
generators=[topo.pattern.Disk(
x=numbergen.UniformRandom(lbound=-2.125,ubound=2.125, seed=p.input_seed+12),
y=numbergen.UniformRandom(lbound=-2.125,ubound=2.125, seed=p.input_seed+36),
size=2.0, aspect_ratio=1.0, scale = disk_scale,
bounds=sheet.BoundingBox(radius=1.125),smoothing=0.1),
topo.pattern.random.UniformRandom(scale=rand_scale)])]
retina_inputs = topo.pattern.Selector(generators=disks_inputs)
if p.retinal_waves == 0:
retina_inputs = combined_inputs
else:
topo.sim.schedule_command(p.retinal_waves, 'topo.sim["Retina"].set_input_generator(combined_inputs, push_existing=False)')
misc.view_patterns(topo.pattern.Selector(generators=disks_inputs), radius=p.area/2.0+0.25+0.375+0.5)
__GCAL_TY__
### Specify weight initialization, response function, and learning function
projection.CFProjection.cf_shape=pattern.Disk(smoothing=0.0)
projection.CFProjection.response_fn=responsefn.optimized.CFPRF_DotProduct_opt()
projection.CFProjection.learning_fn=learningfn.optimized.CFPLF_Hebbian_opt()
projection.CFProjection.weights_output_fns=[transferfn.optimized.CFPOF_DivisiveNormalizeL1_opt()]
projection.SharedWeightCFProjection.response_fn=responsefn.optimized.CFPRF_DotProduct_opt()
__GCAL_TY__
topo.sim['Retina']=sheet.GeneratorSheet(nominal_density=p.retina_density,
input_generator=retina_inputs, period=1.0, phase=0.05,
nominal_bounds=sheet.BoundingBox(radius=p.area/2.0+0.25+0.375+0.5))
The two LGN sheets have center-surround receptive fields and may have contrast-gain control enabled or disabled.
__GCAL_TY__
lgn_surroundg = pattern.Gaussian(size=0.25,aspect_ratio=1.0, output_fns=[transferfn.DivisiveNormalizeL1()])
# LGN has lateral connections for divisive normalization for GCL and GCAL models
for s in ['LGNOn','LGNOff']:
extra_kwargs = dict(tsettle=2,strict_tsettle=1) if p.gain_control else dict(tsettle=0,strict_tsettle=0)
topo.sim[s]=sheet.optimized.LISSOM_Opt(
nominal_density=p.lgn_density,
nominal_bounds=sheet.BoundingBox(radius=p.area/2.0+0.25+0.5),
output_fns=[transferfn.misc.HalfRectify()],
measure_maps=False, **extra_kwargs)
if p.gain_control:
topo.sim.connect(s,s,delay=0.05, name='LateralGC',
dest_port=("Activity"),activity_group=(0.6, DivideWithConstant(c=0.11)),
connection_type=projection.SharedWeightCFProjection,
strength=0.6,weights_generator=lgn_surroundg,
nominal_bounds_template=sheet.BoundingBox(radius=0.25))
__GCAL_TY__
learning_rate = 0.01 if p.homeostasis else 0.0
topo.sim["V1"] = sheet.lissom.LISSOM(nominal_density=p.cortex_density,
tsettle=16, plastic=True,
nominal_bounds=sheet.BoundingBox(radius=p.area/2.0),
output_fns = [HomeostaticResponse(t_init=p.t_init, learning_rate=learning_rate)])
topo.sim["V1"].joint_norm_fn=topo.sheet.optimized.compute_joint_norm_totals_opt
Projections represent synaptic connectivity between different neural populations. In this developmental model, all projections are initializing with some profile for the initial weights which may change over time (via Hebbian learning) if they are set to be plastic
__GCAL_TY__
centerg = pattern.Gaussian(size=0.07385,aspect_ratio=1.0,
output_fns=[transferfn.DivisiveNormalizeL1()])
surroundg = pattern.Gaussian(size=0.29540,aspect_ratio=1.0,
output_fns=[transferfn.DivisiveNormalizeL1()])
on_weights = pattern.Composite(generators=[centerg,surroundg],operator=numpy.subtract)
off_weights = pattern.Composite(generators=[surroundg,centerg],operator=numpy.subtract)
strength_factor = 6.0
topo.sim.connect(
'Retina','LGNOn',delay=0.05,strength=2.33*strength_factor, name='Afferent',
connection_type=projection.SharedWeightCFProjection,
nominal_bounds_template=sheet.BoundingBox(radius=0.375),
weights_generator=on_weights)
topo.sim.connect(
'Retina','LGNOff',delay=0.05,strength=2.33*strength_factor, name='Afferent',
connection_type=projection.SharedWeightCFProjection,
nominal_bounds_template=sheet.BoundingBox(radius=0.375),
weights_generator=off_weights)
"Center surround (difference-of-Gaussian) weights successfully generated"
Here is a look at the weights from the retina to the LGN sheets and the normalization pool within the LGN for contrast-gain control:
misc.view_weights([topo.sim.LGNOn.Afferent, topo.sim.LGNOff.Afferent, topo.sim.LGNOn.LateralGC])
__GCAL_TY__
# Adjust feedforward delays to allow a common measurement protocol with and without gain control.
LGN_V1_delay = 0.05 if p.gain_control else 0.10
topo.sim.connect(
'LGNOn','V1',delay=LGN_V1_delay,strength=p.aff_strength,name='LGNOnAfferent',
dest_port=('Activity','JointNormalize','Afferent'),
connection_type=projection.CFProjection,learning_rate=p.aff_lr,
nominal_bounds_template=sheet.BoundingBox(radius=0.27083),
weights_generator= pattern.random.GaussianCloud(gaussian_size=2*0.27083),
learning_fn=learningfn.optimized.CFPLF_Hebbian_opt())
topo.sim.connect(
'LGNOff','V1',delay=LGN_V1_delay,strength=p.aff_strength,name='LGNOffAfferent',
dest_port=('Activity','JointNormalize','Afferent'),
connection_type=projection.CFProjection,learning_rate=p.aff_lr,
nominal_bounds_template=sheet.BoundingBox(radius=0.27083),
weights_generator= pattern.random.GaussianCloud(gaussian_size=2*0.27083),
learning_fn=learningfn.optimized.CFPLF_Hebbian_opt())
"Afferent GaussianCloud weights successfully generated."
There are two types of lateral weight in the cortical sheet: long-range lateral inhibitory connectivity and shorter-range lateral excitatory connectivity. Note that this is only a subset of the connectivity of V1, but includes the aspects that lead to map development that account for activity bubble formation.
__GCAL_TY__
lateral_excitatory_weights = pattern.Gaussian(aspect_ratio=1.0, size=0.05)
topo.sim.connect(
'V1','V1',delay=0.05,strength=p.exc_strength,name='LateralExcitatory',
connection_type=projection.CFProjection,learning_rate=p.exc_lr,
nominal_bounds_template=sheet.BoundingBox(radius=0.104),
weights_generator=lateral_excitatory_weights)
"Lateral excitatory weights successfully generated"
__GCAL_TY__
lateral_inhibitory_weights = pattern.random.GaussianCloud(gaussian_size=0.15)
topo.sim.connect(
'V1','V1',delay=0.05,strength=-1.0*p.inh_strength,name='LateralInhibitory',
connection_type=projection.CFProjection,learning_rate=p.inh_lr,
nominal_bounds_template=sheet.BoundingBox(radius=0.22917),
weights_generator=lateral_inhibitory_weights)
"Lateral inhibitory weights successfully generated"
Here are the weights from the LGNOn sheet to V1 and the lateral connectivity within V1 (plastic):
misc.view_weights([topo.sim.V1.LGNOnAfferent, topo.sim.V1.LateralInhibitory, topo.sim.V1.LateralExcitatory])
__GCAL_TY__
### Default locations for model editor
topo.sim.grid_layout([[None, 'V1', None],
['LGNOn', None, 'LGNOff'],
[None, 'Retina', None]], xstart=150,item_scale=0.8)
import topo.analysis.featureresponses
topo.analysis.featureresponses.FeatureMaps.selectivity_multiplier=2.0
contrasts = [{'contrast':c} for c in [100, 80, 60, 40, 20, 10]]
topo.analysis.featureresponses.FeatureCurveCommand.curve_parameters= contrasts
Export all code cells marked with the __GCAL_TY__ identifier will be exported to gcal.ty
from topo.misc.ipython import export_notebook
diff = export_notebook('./gcal.html', 'gcal.ty', identifier='__GCAL_TY__')
from topo.misc.ipython import RunProgress
RunProgress().run(3000)
Looking at the new values for the weights below, you can see that,
misc.view_weights([topo.sim.V1.LGNOnAfferent, topo.sim.V1.LateralInhibitory, topo.sim.V1.LateralExcitatory])
Looking at the sample activity patterns below,
misc.view_activities([topo.sim.Retina, topo.sim.LGNOn, topo.sim.LGNOff, topo.sim.V1])
Present 24 sine gratings and copute the average preferred orientation for each V1 neuron:
from topo.command.analysis import measure_sine_pref
measure_sine_pref()['V1']['OrientationPreference'].top