%% Force Method with Izhikevich Network
clear all
close all
clc
T = 15000; %Total time in ms
dt = 0.04; %Integration time step in ms
nt = round(T/dt); %Time steps
N = 2000; %Number of neurons
%% Izhikevich Parameters
C = 250; %capacitance
vr = -60; %resting membrane
b = -2; %resonance parameter
ff = 2.5; %k parameter for Izhikevich, gain on v
vpeak = 30; % peak voltage
vreset = -65; % reset voltage
vt = vr+40-(b/ff); %threshold %threshold
u = zeros(N,1); %initialize adaptation
a = 0.01; %adaptation reciprocal time constant
d = 200; %adaptation jump current
tr = 2; %synaptic rise time
td = 20; %decay time
p = 0.1; %sparsity
G =5*10^3; %Gain on the static matrix with 1/sqrt(N) scaling weights. Note that the units of this have to be in pA.
Q =5*10^3; %Gain on the rank-k perturbation modified by RLS. Note that the units of this have to be in pA
Irh = 0.25*ff*(vt-vr)^2;
%Storage variables for synapse integration
IPSC = zeros(N,1); %post synaptic current
h = zeros(N,1);
r = zeros(N,1);
hr = zeros(N,1);
JD = zeros(N,1);
%-----Initialization---------------------------------------------
v = vr+(vpeak-vr)*rand(N,1); %initial distribution
v_ = v; %These are just used for Euler integration, previous time step storage
rng(1)
%% Target signal COMMENT OUT TEACHER YOU DONT WANT, COMMENT IN TEACHER YOU WANT.
zx = (sin(2*5*pi*(1:1:nt)*dt/1000));
%%
k = min(size(zx)); %used to get the dimensionality of the approximant correctly. Typically will be 1 unless you specify a k-dimensional target function.
OMEGA = G*(randn(N,N)).*(rand(N,N)<p)/(p*sqrt(N)); %Static weight matrix.
z = zeros(k,1); %initial approximant
BPhi = zeros(N,k); %initial decoder. Best to keep it at 0.
tspike = zeros(5*nt,2); %If you want to store spike times,
ns = 0; %count toal number of spikes
BIAS = 1000; %Bias current, note that the Rheobase is around 950 or something. I forget the exact formula for this but you can test it out by shutting weights and feeding constant currents to neurons
E = (2*rand(N,k)-1)*Q; %Weight matrix is OMEGA0 + E*BPhi';
%%
Pinv = eye(N)*2; %initial correlation matrix, coefficient is the regularization constant as well
step = 20; %optimize with RLS only every 50 steps
imin = round(5000/dt); %time before starting RLS, gets the network to chaotic attractor
icrit = round(10000/dt); %end simulation at this time step
current = zeros(nt,k); %store the approximant
RECB = zeros(nt,5); %store the decoders
REC = zeros(nt,10); %Store voltage and adaptation variables for plotting
i=1;
%% SIMULATION
tic
ilast = i ;
%icrit = ilast; %uncomment this, and restart cell if you want to test
% performance before icrit.
for i = ilast:1:nt;
%% EULER INTEGRATE
I = IPSC + E*z + BIAS; %postsynaptic current
v = v + dt*(( ff.*(v-vr).*(v-vt) - u + I))/C ; % v(t) = v(t-1)+dt*v'(t-1)
u = u + dt*(a*(b*(v_-vr)-u)); %same with u, the v_ term makes it so that the integration of u uses v(t-1), instead of the updated v(t)
%%
index = find(v>=vpeak);
if length(index)>0
JD = sum(OMEGA(:,index),2); %compute the increase in current due to spiking
tspike(ns+1:ns+length(index),:) = [index,0*index+dt*i]; %uncomment this
%if you want to store spike times. Takes longer.
ns = ns + length(index);
end
%synapse for single exponential
if tr == 0
IPSC = IPSC*exp(-dt/td)+ JD*(length(index)>0)/(td);
r = r *exp(-dt/td) + (v>=vpeak)/td;
else
%synapse for double exponential
IPSC = IPSC*exp(-dt/tr) + h*dt;
h = h*exp(-dt/td) + JD*(length(index)>0)/(tr*td); %Integrate the current
r = r*exp(-dt/tr) + hr*dt;
hr = hr*exp(-dt/td) + (v>=vpeak)/(tr*td);
end
z = BPhi'*r; %approximant
err = z - zx(:,i); %error
%% RLS
if mod(i,step)==1
if i > imin
if i < icrit
cd = Pinv*r;
BPhi = BPhi - (cd*err');
Pinv = Pinv -((cd)*(cd'))/( 1 + (r')*(cd));
end
end
end
%% Store, and plot.
u = u + d*(v>=vpeak); %implements set u to u+d if v>vpeak, component by component.
v = v+(vreset-v).*(v>=vpeak); %implements v = c if v>vpeak add 0 if false, add c-v if true, v+c-v = c
v_ = v; % sets v(t-1) = v for the next itteration of loop
REC(i,:) = [v(1:5)',u(1:5)'];
current(i,:) = z';
RECB(i,:)=BPhi(1:5);
if mod(i,round(100/dt))==1
drawnow
gg = max(1,i - round(3000/dt)); %only plot for last 3 seconds
figure(2)
plot(dt*(gg:1:i)/1000,zx(:,gg:1:i),'k','LineWidth',2), hold on
plot(dt*(gg:1:i)/1000,current(gg:1:i,:),'b--','LineWidth',2), hold off
xlabel('Time (s)')
ylabel('$\hat{x}(t)$','Interpreter','LaTeX')
legend('Approximant','Target Signal')
xlim([dt*i-3000,dt*i]/1000)
figure(3)
plot((1:1:i)*dt/1000,RECB(1:1:i,:))
figure(14)
plot(tspike(1:ns,2),tspike(1:ns,1),'k.')
ylim([0,100])
end
end
%%
tspike = tspike(tspike(:,2)~=0,:);
M = tspike(tspike(:,2)>dt*icrit);
AverageFiringRate = 1000*length(M)/(N*(T-dt*icrit))
%% Plotting neurons before and after learning
figure(30)
for j = 1:1:5
plot((1:1:i)*dt/1000,REC(1:1:i,j)/(vpeak-vreset)+j), hold on
end
xlim([T/1000-2,T/1000])
xlabel('Time (s)')
ylabel('Neuron Index')
title('Post Learning')
figure(31)
for j = 1:1:5
plot((1:1:i)*dt/1000,REC(1:1:i,j)/(vpeak-vreset)+j), hold on
end
xlim([0,imin*dt/1000])
xlabel('Time (s)')
ylabel('Neuron Index')
title('Pre-Learning')
figure(40)
Z = eig(OMEGA+E*BPhi'); %eigenvalues after learning
Z2 = eig(OMEGA); %eigenvalues before learning
%%
plot(Z2,'r.'), hold on
plot(Z,'k.')
legend('Pre-Learning','Post-Learning')
xlabel('Re \lambda')
ylabel('Im \lambda')