% Fuhs and Touretzky (2006)'s developmental model
% eric zilli - 20111214 - v1.0
%
% This script demonstrates Fuhs and Touretzky (2006)'s developmental model
% of the formation of symmetric rings of output from a cell onto the other
% cells in the rectangular sheet of cells.
%
% In the model, sinusoidal gratings called "wave packets" flow across the
% sheet of cells, each beginning with a random orientationg and moving
% from one side to the other, in a direction perpendicular to the
% orientation of its stripes.
%
% The BCM-like learning rule associates a cell to the cells coactive with
% it, so as the wave packets pass any individual cell, the sinusoidal
% grating gets lightly stamped into its outputs. As gratings of random
% orientations keep passing over it, its synaptic weights contain the
% combination of gratings at every orientation, which eventually produces
% symmetric rings of synaptic output.
%
% The desired ring pattern should start to appear once 20-30 packets
% have passed (depending on the random orientations they had).
%
% This code is released into the public domain. Not for use in skynet.
% if >0, plots the sheet of activity and weights during the simulation on
% every livePlot'th step
livePlot = 100;
% Weight matrix options
useCurrentW = 0; % if W exists in namespace, use that one instead of loading/generating
loadWIfPossible = 0;
saveW = 0; % save W to disk after generating
%% Network/Weight matrix parameters
n = 62;
ncells = n*n; % total number of cells in network
% frequency of wave packets
kappa = 9*pi/31;
wavePerPacket = 3;
% tonic firing rate of cells during development (wave packets push activity
% from 0 to 2)
ftonic = 1;
%% Simulation parameters
dt = 1; % time step, ms
npackets = 1000;
simdur = 1e3; % time per packet simulation, ms
t = 0; % simulation time variable, ms
tind = 0; % time step number
x = 0; % position, cm
y = 0; % position, cm
%% Initial conditions
f = zeros(1,ncells); % initial firing rate
%% Make x a 2-by-ncells vector of the 2D cell positions on the neural sheet
% plot as: figure; imagesc(reshape(cellDists,n,n))
x = linspace(-round(n/2),round(n/2),n);
[X,Y] = meshgrid(x,x);
x = [reshape(X,1,[]); reshape(Y,1,[])];
cellDists = sqrt(x(1,:).^2 + x(2,:).^2); % distance from (0,0)
% convert to polar coordinates:
% get angles of cells around (0,0)
theta = atan2(x(2,:),x(1,:));
%% Optionally load weight matrix
fname = sprintf('data/W_FT2006_dev_n%d.mat',ncells);
if ~(useCurrentW && exist('W','var'))
if loadWIfPossible && exist(fname,'file')
fprintf('Attempting to load pre-generated W...\n')
load(fname);
fprintf('+ Loaded pre-generated W.\n')
else
W = zeros(ncells);
end
end
%% Make optional figure of sheet of activity
if livePlot
h = figure('color','w','name','Activity of sheet of cells on brain''s surface');
set(h,'position',[520 378 1044 420])
drawnow
end
%% !! Learn weight matrix (main loop)
fprintf('Development starting. Press ctrl+c to end...\n')
for packet=1:npackets
t = 0;
tind = -40;
% random direction for this packet:
packetDir = 2*pi*rand;
% learning rate decreases from 2e3 to 1e5 "units?"
% but they don't really tell us how. I'm guessing they mean
% it slowed exponentially toward 1e5 rather than growing exponentially
% toward it, and since we don't know the growth rate, I'll guess:
tauw = 2e3 + (1e5-2e3)*(2./(1+exp(-packet/200))-1);
%% Time within each packet
while t2*pi*wavePerPacket) = ftonic;
% end this run if the packet has left the cells
if all(phases>0) && all(all(f==ftonic))
break
end
% update weight matrix
% we use their approximation that mean(f) = ftonic
W = W + dt/tauw*((f' - ftonic)*f - W);
if livePlot>0 && (livePlot==1 || mod(tind,livePlot)==1)
figure(h);
subplot(121);
image(32*reshape(f,n,n));
title(sprintf('Activity of sheet of cells; t = %.1f ms, packet %d',t,packet))
axis square
set(gca,'ydir','normal')
subplot(122);
imagesc(reshape(W(:,ncells/2-round(n/2)),n,n));
axis square
set(gca,'ydir','normal')
title({'Weights from center cell onto all other cells'})
drawnow
end
end
end
%% Optionally save weight matrix
if saveW
save(fname,'W','-v7.3');
end