Add files via upload

jocksanv · Aug 26, 2017 · 7856edb · 7856edb
1 parent 8543d27
commit 7856edb
Show file tree

Hide file tree

Showing 44 changed files with 235 additions and 0 deletions.
diff --git a/exercises/c1/c1.pdf b/exercises/c1/c1.pdf
diff --git a/exercises/c1/c1.ps.gz b/exercises/c1/c1.ps.gz
diff --git a/exercises/c1/c1cd.tar.gz b/exercises/c1/c1cd.tar.gz
diff --git a/exercises/c1/data/c10p1.mat b/exercises/c1/data/c10p1.mat
@@ -0,0 +1,100 @@
+   1.2610673e+00   8.7850842e-02
+   9.4870162e-01  -1.8085079e-02
+   8.6185491e-01   1.6224918e-01
+   9.5081336e-01  -4.3257720e-01
+   1.3874176e+00  -7.7076637e-02
+   1.4257554e+00  -2.1140316e-01
+   1.0510257e+00  -2.3929085e-01
+   8.6430210e-01   2.0747363e-01
+   1.0557558e+00  -4.4090904e-02
+   9.8515345e-01  -4.5922628e-01
+   1.2154257e+00   1.3510720e-01
+   1.2889587e+00   2.2091317e-01
+   1.3419986e+00  -1.2473362e-01
+   6.6064156e-01   2.8451805e-01
+   1.2035246e+00   8.0300324e-01
+   8.7489386e-01   4.1143334e-01
+   1.3307248e+00  -2.2338402e-01
+   9.9965170e-01  -2.7229713e-01
+   8.0970832e-01  -1.8176388e-01
+   1.1561395e+00   3.1659399e-01
+   9.6945682e-01  -1.8446872e-02
+   1.1501202e+00   1.1455769e-01
+   8.9451168e-01   3.9948466e-01
+   7.4866926e-01   6.2512980e-03
+   1.4240451e+00  -5.1108909e-02
+   9.5394331e-01   3.5143469e-01
+   8.5480809e-01  -1.4734469e-01
+   9.3359112e-01  -9.1376048e-02
+   9.3737253e-01   9.2936169e-02
+   7.7042668e-01  -7.7022002e-02
+   1.0234778e+00  -2.9261726e-01
+   1.3611381e+00   2.0925376e-01
+   6.5553227e-01  -4.0112295e-01
+   9.1105658e-01  -4.8095323e-01
+   1.1521568e+00  -2.3549257e-01
+   8.6928331e-01   4.8344866e-02
+   1.1996496e+00   5.3552165e-01
+   8.5463939e-01   9.0363593e-02
+   1.0673105e+00  -2.2429837e-02
+   1.3079029e+00   6.1442130e-01
+   1.1532697e+00   2.4437677e-01
+   5.1063739e-01   5.5802326e-02
+   1.1043312e+00   1.0781936e-01
+   1.2285890e+00  -4.1432114e-02
+   6.3112986e-01   6.9896014e-02
+   9.1630523e-01  -2.2843712e-01
+   9.3854954e-01   3.8855984e-02
+   1.1032298e+00  -1.4304489e-01
+   1.0082623e+00   8.2076036e-02
+   1.3657724e+00  -9.5120396e-02
+   1.6112417e-01   9.3146135e-01
+   1.2744314e-02   7.1172188e-01
+  -1.0664676e-01   1.3752386e+00
+  -8.9814343e-02   6.2681146e-01
+  -4.4419760e-01   6.8385513e-01
+  -2.0725695e-01   7.0481888e-01
+   2.3524013e-02   1.0151033e+00
+   1.3890694e-01   8.7296951e-01
+  -3.1440981e-01   1.0613184e+00
+  -1.3083182e-01   1.3129298e+00
+   1.3539111e-01   1.1399145e+00
+  -9.8466091e-02   1.3231128e+00
+  -4.4501379e-01   8.2839910e-01
+   5.9267664e-01   3.5780148e-01
+   2.5489893e-01   6.6979484e-01
+  -2.0199882e-01   1.0335235e+00
+  -4.2885576e-01   1.0911313e+00
+   1.0322712e-01   5.9400205e-01
+  -1.9723433e-01   1.3403960e+00
+  -1.1353115e-01   6.7319187e-01
+   4.0688787e-01   9.8696848e-01
+  -4.6796045e-02   6.6406917e-01
+   3.2124189e-01   1.3173497e+00
+   5.1243938e-03   9.1821940e-01
+   2.8407226e-01   1.0560155e+00
+   9.4868475e-02   9.3456735e-01
+  -9.5926837e-02   1.0247360e+00
+   4.0200412e-01   1.2523789e+00
+   2.4620030e-01   1.0425304e+00
+   2.0865607e-01   1.5049251e+00
+  -2.8119147e-01   1.2229519e+00
+   1.1669107e-01   1.2967161e+00
+   1.2690337e-01   9.1158256e-01
+   2.2537798e-01   1.1935403e+00
+  -9.9160040e-02   1.1949584e+00
+  -1.6107908e-01   8.4973897e-01
+   2.6860421e-01   9.4216399e-01
+   3.0917642e-01   9.6035139e-01
+   1.9981889e-03   1.1045532e+00
+  -1.7998092e-01   6.0799009e-01
+   4.0607064e-01   1.1784296e+00
+   2.2849174e-01   1.3944892e+00
+  -2.7332032e-01   1.3102919e+00
+  -4.0823356e-01   1.1200681e+00
+  -2.0742396e-01   1.3783053e+00
+  -1.6515944e-01   1.0489970e+00
+  -4.4035092e-01   1.6121930e+00
+   3.0303782e-01   1.0023180e+00
+   1.1252761e-02   5.4663902e-01
+  -4.0678274e-01   1.0435023e+00
diff --git a/exercises/c1/data/c1p8.mat b/exercises/c1/data/c1p8.mat
diff --git a/exercises/c10/c10.pdf b/exercises/c10/c10.pdf
diff --git a/exercises/c10/c10.ps.gz b/exercises/c10/c10.ps.gz
diff --git a/exercises/c10/c10cd.tar.gz b/exercises/c10/c10cd.tar.gz
diff --git a/exercises/c10/c10code.tar.gz b/exercises/c10/c10code.tar.gz
diff --git a/exercises/c10/code/c10p10.m b/exercises/c10/code/c10p10.m
@@ -0,0 +1 @@
+% c10p10.m%% demonstration of the Helmholtz machine using the wake-sleep% algorithm. rand('state',105);ndim=4;        % dimension of the square of barsndimsq=ndim^2;nu=400;        % number of examplespbar=0.4;      % probability that a given bar is onpout=0.9;      % extra noise added at the endii=ones(1,ndim);oo=zeros(1,ndim);conv=zeros(ndimsq,ndim);for i=1:ndim  conv((i-1)*ndim+(1:ndim),i)=ones(ndim,1);end	       bars=conv*(rand(ndim,nu)<pbar);  % which bars are on in which examplesu=bars.*(rand(ndimsq,nu)<pout) + (1-bars).*(rand(ndimsq,nu)>pout);                                 % generate output noisenhid=ndim+1;   % number of hidden unitsW=zeros(nhid,ndimsq);G=zeros(ndimsq,nhid);g=zeros(nhid,1);    % generative biases for vh=zeros(ndimsq,1);  % generative biases for uw=zeros(nhid,1);    % recognition biases for vsig=inline('1./(1+exp(-x))');  % sigmoid function				 mini=20;            % mini-batches of 20 input cases at a timenits=100000;eta=0.1/mini;       % learning ratefor it=1:nits  %wake phase  rp=randperm(nu);    au=u(:,rp(1:mini));   % choose a random set of mini patterns  v = rand(nhid,mini) < sig(W*au + repmat(w,1,mini));    % propagate u->v and sample  gv = sig(g);          % top down prediction of v  g = g + eta*(v - repmat(gv,1,mini))*ones(mini,1);                         % delta rule for g  gpu = sig(G*v + repmat(h,1,mini));   % top down prediction of u  h = h + eta*(au - gpu)*ones(mini,1); % delta rule for h  G = G + eta*(au - gpu)*v';           % delta rule for G  sv = v;                              % for plotting  %sleep phase  gv = sig(g);          % top down prediction of v  v = rand(nhid,mini)<repmat(gv,1,mini); % top down sample of v  gu = rand(ndimsq,mini) < sig(G*v + repmat(h,1,mini)); % top down sample of u  rv = sig(W*gu + repmat(w,1,mini));  w = w + eta*(v - rv)*ones(mini,1);   % delta rule for w  W = W + eta*(v - rv)*gu';            % delta rule for W  if (mod(it,200)==0)    disp(sprintf('%d',it));    clf;    colormap(gray);    for i=1:nhid      subplot(4,nhid,i);      imagesc(reshape(G(:,i),ndim,ndim),[min(G(:)) max(G(:))]);      set(gca,'xtick',[],'ytick',[]);      title(num2str(sig(g(i))));       % title is gen prob      colorbar;      subplot(4,nhid,i+nhid);      imagesc(reshape(W(i,:),ndim,ndim),[min(W(:)) max(W(:))]);      set(gca,'xtick',[],'ytick',[]);      colorbar;      subplot(4,nhid,i+2*nhid);      imagesc(reshape(u(:,rp(i)),ndim,ndim),[0 1]);      set(gca,'xtick',[],'ytick',[]);      title(rp(i));                    % title is pattern number      colorbar;      subplot(4,nhid,i+3*nhid);      imagesc(reshape(gpu(:,i),ndim,ndim),[0 1]);      set(gca,'xtick',[],'ytick',[]);        title(sprintf('%1d',sv(:,i)));   % title is hidden unit state      colorbar;    end    drawnow;  endend      

diff --git a/exercises/c10/code/c10p11.m b/exercises/c10/code/c10p11.m
@@ -0,0 +1 @@
+% c10p11.m%% demonstration of the Helmholtz machine using the wake-sleep% algorithm. rand('state',105);ndim=4;        % dimension of the square of barsndimsq=ndim^2;nu=400;        % number of examplespbar=0.5;      % probability that a given bar is onpout=0.05;      % extra noise added at the endii=ones(1,ndim);oo=zeros(1,ndim);conv=zeros(ndimsq,ndim);for i=1:ndim  conv((i-1)*ndim+(1:ndim),i)=ones(ndim,1);end	       bars=conv*([rand(ndim/2,nu/2)<pbar zeros(ndim/2,nu/2); ...      zeros(ndim/2,nu/2) rand(ndim/2,nu/2)<pbar]);  % which bars are on in which examplesu=bars.*(rand(ndimsq,nu)>pout) + (1-bars).*(rand(ndimsq,nu)<pout);                                 % generate output noise

diff --git a/exercises/c10/code/c10p6.m b/exercises/c10/code/c10p6.m
@@ -0,0 +1 @@
+% c10p6.m%% factor analysis exercisenu=500;   % number of data pointsrandn('state',100); % to make the same as in the textgtrue=[1 ; 1 ; 1];  % since the underlying factor is identical in                    % each observable componentvtrue=randn(1,nu);      u1=gtrue*vtrue + 0.5*randn(3,nu); % data for 10.4A and Bu2=u1;u2(2,:)=vtrue+3*randn(1,nu);      % corrupt u_2 by more noiseC1=cov(u1');        % covariance matrix for u1C2=cov(u2');        % covariance matrix for u2                    % the means are not very important since there                    % are so many data points.rot=[cos(pi/3) sin(pi/3) 0 ; -sin(pi/3) cos(pi/3) 0 ; 0 0 1] * ...    [cos(pi/6) 0 sin(pi/6) ; 0 1 0 ; -sin(pi/6) 0 cos(pi/6)];

diff --git a/exercises/c10/data/c10p1.mat b/exercises/c10/data/c10p1.mat
@@ -0,0 +1,100 @@
+   1.2610673e+00   8.7850842e-02
+   9.4870162e-01  -1.8085079e-02
+   8.6185491e-01   1.6224918e-01
+   9.5081336e-01  -4.3257720e-01
+   1.3874176e+00  -7.7076637e-02
+   1.4257554e+00  -2.1140316e-01
+   1.0510257e+00  -2.3929085e-01
+   8.6430210e-01   2.0747363e-01
+   1.0557558e+00  -4.4090904e-02
+   9.8515345e-01  -4.5922628e-01
+   1.2154257e+00   1.3510720e-01
+   1.2889587e+00   2.2091317e-01
+   1.3419986e+00  -1.2473362e-01
+   6.6064156e-01   2.8451805e-01
+   1.2035246e+00   8.0300324e-01
+   8.7489386e-01   4.1143334e-01
+   1.3307248e+00  -2.2338402e-01
+   9.9965170e-01  -2.7229713e-01
+   8.0970832e-01  -1.8176388e-01
+   1.1561395e+00   3.1659399e-01
+   9.6945682e-01  -1.8446872e-02
+   1.1501202e+00   1.1455769e-01
+   8.9451168e-01   3.9948466e-01
+   7.4866926e-01   6.2512980e-03
+   1.4240451e+00  -5.1108909e-02
+   9.5394331e-01   3.5143469e-01
+   8.5480809e-01  -1.4734469e-01
+   9.3359112e-01  -9.1376048e-02
+   9.3737253e-01   9.2936169e-02
+   7.7042668e-01  -7.7022002e-02
+   1.0234778e+00  -2.9261726e-01
+   1.3611381e+00   2.0925376e-01
+   6.5553227e-01  -4.0112295e-01
+   9.1105658e-01  -4.8095323e-01
+   1.1521568e+00  -2.3549257e-01
+   8.6928331e-01   4.8344866e-02
+   1.1996496e+00   5.3552165e-01
+   8.5463939e-01   9.0363593e-02
+   1.0673105e+00  -2.2429837e-02
+   1.3079029e+00   6.1442130e-01
+   1.1532697e+00   2.4437677e-01
+   5.1063739e-01   5.5802326e-02
+   1.1043312e+00   1.0781936e-01
+   1.2285890e+00  -4.1432114e-02
+   6.3112986e-01   6.9896014e-02
+   9.1630523e-01  -2.2843712e-01
+   9.3854954e-01   3.8855984e-02
+   1.1032298e+00  -1.4304489e-01
+   1.0082623e+00   8.2076036e-02
+   1.3657724e+00  -9.5120396e-02
+   1.6112417e-01   9.3146135e-01
+   1.2744314e-02   7.1172188e-01
+  -1.0664676e-01   1.3752386e+00
+  -8.9814343e-02   6.2681146e-01
+  -4.4419760e-01   6.8385513e-01
+  -2.0725695e-01   7.0481888e-01
+   2.3524013e-02   1.0151033e+00
+   1.3890694e-01   8.7296951e-01
+  -3.1440981e-01   1.0613184e+00
+  -1.3083182e-01   1.3129298e+00
+   1.3539111e-01   1.1399145e+00
+  -9.8466091e-02   1.3231128e+00
+  -4.4501379e-01   8.2839910e-01
+   5.9267664e-01   3.5780148e-01
+   2.5489893e-01   6.6979484e-01
+  -2.0199882e-01   1.0335235e+00
+  -4.2885576e-01   1.0911313e+00
+   1.0322712e-01   5.9400205e-01
+  -1.9723433e-01   1.3403960e+00
+  -1.1353115e-01   6.7319187e-01
+   4.0688787e-01   9.8696848e-01
+  -4.6796045e-02   6.6406917e-01
+   3.2124189e-01   1.3173497e+00
+   5.1243938e-03   9.1821940e-01
+   2.8407226e-01   1.0560155e+00
+   9.4868475e-02   9.3456735e-01
+  -9.5926837e-02   1.0247360e+00
+   4.0200412e-01   1.2523789e+00
+   2.4620030e-01   1.0425304e+00
+   2.0865607e-01   1.5049251e+00
+  -2.8119147e-01   1.2229519e+00
+   1.1669107e-01   1.2967161e+00
+   1.2690337e-01   9.1158256e-01
+   2.2537798e-01   1.1935403e+00
+  -9.9160040e-02   1.1949584e+00
+  -1.6107908e-01   8.4973897e-01
+   2.6860421e-01   9.4216399e-01
+   3.0917642e-01   9.6035139e-01
+   1.9981889e-03   1.1045532e+00
+  -1.7998092e-01   6.0799009e-01
+   4.0607064e-01   1.1784296e+00
+   2.2849174e-01   1.3944892e+00
+  -2.7332032e-01   1.3102919e+00
+  -4.0823356e-01   1.1200681e+00
+  -2.0742396e-01   1.3783053e+00
+  -1.6515944e-01   1.0489970e+00
+  -4.4035092e-01   1.6121930e+00
+   3.0303782e-01   1.0023180e+00
+   1.1252761e-02   5.4663902e-01
+  -4.0678274e-01   1.0435023e+00
diff --git a/exercises/c2/c2.pdf b/exercises/c2/c2.pdf
diff --git a/exercises/c2/c2.ps.gz b/exercises/c2/c2.ps.gz
diff --git a/exercises/c2/c2cd.tar.gz b/exercises/c2/c2cd.tar.gz
diff --git a/exercises/c2/c2code.tar.gz b/exercises/c2/c2code.tar.gz
diff --git a/exercises/c2/code/c2p5.m b/exercises/c2/code/c2p5.m
@@ -0,0 +1 @@
+% mapping from cortex into the retinalam=1.2   % constants for the mapep0=1.0;[X,Y]=meshgrid(0.05:.01:3,-2:.01:2);             % global coordinates for the cortexep=ep0*(exp(X/lam)-1);              % eccentricitya=-(180*(ep0+ep).*Y)./(lam*ep*pi);  % azimuthin=abs(a)<90;                       a=a.*(abs(a)<90)+90*(a>=90) - 90*(a<=-90);                                    % take care of absurditiesclf;K=8;Theta=pi/6;Phi=0;s=cos(K*X*cos(Theta)+K*Y*sin(Theta)-Phi);subplot(2,2,1);imagesc(X(1,:),Y(:,1),s.*in);set(gca,'fontsize',18,'fontname','palatino');title('visual cortex');xlabel('X (cm)');ylabel('Y (cm)');daspect([1 1.5 1]);subplot(2,2,2);contourf(ep.*cos(a*2*pi/360),ep.*sin(a*2*pi/360),s,'w:');set(gca,'xlim',[0 12],'ylim',[-12 12]);set(gca,'fontsize',18,'fontname','palatino');title('visual space');daspect([1 1 1]);xlabel('degrees');ylabel('degrees');

diff --git a/exercises/c2/data/c1p8.mat b/exercises/c2/data/c1p8.mat
diff --git a/exercises/c2/data/c2p3.mat b/exercises/c2/data/c2p3.mat
diff --git a/exercises/c3/c3.pdf b/exercises/c3/c3.pdf
diff --git a/exercises/c3/c3.ps.gz b/exercises/c3/c3.ps.gz
diff --git a/exercises/c3/c3cd.tar.gz b/exercises/c3/c3cd.tar.gz
diff --git a/exercises/c3/c3code.tar.gz b/exercises/c3/c3code.tar.gz
diff --git a/exercises/c3/code/c3p10.m b/exercises/c3/code/c3p10.m
@@ -0,0 +1 @@
+function [est,K,ind]=c3p10(stim,spk,nfft)%[est,K,sstim]=wf(stim,spk,nfft)%Finds the optimal decoding filter K to predict stim from spk using %matlab function spectrum with nfft frequencies%est    is the resulting stimulus estimate%K      is the filter%ind    is the indices such that est approximmates stim(ind).spk=spk-mean(spk);        % remove the mean of the spikesmstim=mean(stim);         stim=stim-mstim;          % remove the mean of the stimulusP=spectrum(spk,stim,nfft); % work out the various self- and                           % cross-spectral termsTxy=P(:,4);                % the complex transfer function from X to Y --                           % ie the FFT of the decoding filterT=[Txy ; conj(Txy((end-1):-1:2))]; % want the real-valued versionK=fftshift(real(ifft(T)));% shift to make it centeredest=fftfilt(K,spk);       % use the fast filter algorithmest=est(nfft+1:end)+mstim; % get rid of the transientind=(nfft/2+1):(length(stim)-nfft/2); % the indices of stim that                                      % match those of est                                       					   

diff --git a/exercises/c4/c4.pdf b/exercises/c4/c4.pdf
diff --git a/exercises/c4/c4.ps.gz b/exercises/c4/c4.ps.gz
diff --git a/exercises/c5/c5.pdf b/exercises/c5/c5.pdf
diff --git a/exercises/c5/c5.ps.gz b/exercises/c5/c5.ps.gz
diff --git a/exercises/c6/c6.pdf b/exercises/c6/c6.pdf
diff --git a/exercises/c6/c6.ps.gz b/exercises/c6/c6.ps.gz
diff --git a/exercises/c7/c7.pdf b/exercises/c7/c7.pdf
diff --git a/exercises/c7/c7.ps.gz b/exercises/c7/c7.ps.gz
diff --git a/exercises/c7/c7cd.tar.gz b/exercises/c7/c7cd.tar.gz
diff --git a/exercises/c7/c7code.tar.gz b/exercises/c7/c7code.tar.gz
diff --git a/exercises/c7/code/c7p5.m b/exercises/c7/code/c7p5.m
@@ -0,0 +1 @@
+% c7p5.m % % symmetric recurrent nonlinear networkM=[0 -1 ; -1 0];  % connection matrixh=[1 ; 1];        % inputiters=100;        % maximum number of iterationseta=0.1;          % Euler stepclf;subplot(1,2,1);hold on;beta=2.0;         % high gain for the activation function                  % now step the starting conditions around the outside                  % of a grid v1=0;for v2=0:.25:3  c7p5sub(M,beta,h,[v1 ; v2],eta,iters);  drawnow;endv2=0;for v1=0:.25:3  c7p5sub(M,beta,h,[v1 ; v2],eta,iters);  drawnow;endv1=3;for v2=0:.25:3  c7p5sub(M,beta,h,[v1 ; v2],eta,iters);  drawnow;endv2=3;for v1=0:.25:3  c7p5sub(M,beta,h,[v1 ; v2],eta,iters);  drawnow;end                  % now label the graphaxis('square'); grid;title(sprintf('beta=%f',beta),'FontSize',20);xlabel('v_1','FontSize',16); ylabel('v_2','FontSize',16);subplot(1,2,2);hold on;beta=0.5;         % low gain for the activation function                  % now step the starting conditions around the outside                  % of a grid v1=0;for v2=0:.25:3  c7p5sub(M,beta,h,[v1 ; v2],eta,iters);  drawnow;endv2=0;for v1=0:.25:3  c7p5sub(M,beta,h,[v1 ; v2],eta,iters);  drawnow;endv1=3;for v2=0:.25:3  c7p5sub(M,beta,h,[v1 ; v2],eta,iters);  drawnow;endv2=3;for v1=0:.25:3  c7p5sub(M,beta,h,[v1 ; v2],eta,iters);  drawnow;endaxis('square'); grid;title(sprintf('beta=%f',beta),'FontSize',20);xlabel('v_1','FontSize',16); ylabel('v_2','FontSize',16);

diff --git a/exercises/c7/code/c7p5sub.m b/exercises/c7/code/c7p5sub.m
@@ -0,0 +1,28 @@
+function vtrace = c7p1sub(M,beta,h,vinit,eta,iters)
+%vtrace = c7p1sub(M,beta,h,vinit,eta,iters)
+%
+%This function simulates a Hopfield network using simple Euler integration.
+%
+% M     is the connection matrix (should be symmetric)
+% beta  is the gain of the activation function (tanh by befault)
+% h     is the bias
+% vinit is the initial condition vector for the v's
+% eta   is the timestep for the Euler integration
+% iters is the number of iterations to simulate
+%
+% vtrace   is the v vector state of the network over time, 
+%          one column per timestep in the simulation
+
+N = size(M,1);
+vtrace = zeros(N,iters);
+
+vtrace(:,1) = vinit(:);
+
+for i=2:iters
+ vtrace(:,i)=vtrace(:,i-1) + eta*(-vtrace(:,i-1) + ...
+     beta*max(M*vtrace(:,i-1) + h,0));
+end
+
+plot(vtrace(1,:),vtrace(2,:),'o',vtrace(1,:),vtrace(2,:));
+
+
diff --git a/exercises/c8/c8.pdf b/exercises/c8/c8.pdf
diff --git a/exercises/c8/c8.ps.gz b/exercises/c8/c8.ps.gz
diff --git a/exercises/c9/c9.pdf b/exercises/c9/c9.pdf
diff --git a/exercises/c9/c9.ps.gz b/exercises/c9/c9.ps.gz
diff --git a/exercises/call.pdf b/exercises/call.pdf
diff --git a/exercises/call.ps.gz b/exercises/call.ps.gz
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		% c10p10.m%% demonstration of the Helmholtz machine using the wake-sleep% algorithm. rand('state',105);ndim=4; % dimension of the square of barsndimsq=ndim^2;nu=400; % number of examplespbar=0.4; % probability that a given bar is onpout=0.9; % extra noise added at the endii=ones(1,ndim);oo=zeros(1,ndim);conv=zeros(ndimsq,ndim);for i=1:ndim conv((i-1)ndim+(1:ndim),i)=ones(ndim,1);end bars=conv(rand(ndim,nu)<pbar); % which bars are on in which examplesu=bars.(rand(ndimsq,nu)<pout) + (1-bars).(rand(ndimsq,nu)>pout); % generate output noisenhid=ndim+1; % number of hidden unitsW=zeros(nhid,ndimsq);G=zeros(ndimsq,nhid);g=zeros(nhid,1); % generative biases for vh=zeros(ndimsq,1); % generative biases for uw=zeros(nhid,1); % recognition biases for vsig=inline('1./(1+exp(-x))'); % sigmoid function mini=20; % mini-batches of 20 input cases at a timenits=100000;eta=0.1/mini; % learning ratefor it=1:nits %wake phase rp=randperm(nu); au=u(:,rp(1:mini)); % choose a random set of mini patterns v = rand(nhid,mini) < sig(Wau + repmat(w,1,mini)); % propagate u->v and sample gv = sig(g); % top down prediction of v g = g + eta(v - repmat(gv,1,mini))ones(mini,1); % delta rule for g gpu = sig(Gv + repmat(h,1,mini)); % top down prediction of u h = h + eta(au - gpu)ones(mini,1); % delta rule for h G = G + eta(au - gpu)v'; % delta rule for G sv = v; % for plotting %sleep phase gv = sig(g); % top down prediction of v v = rand(nhid,mini)<repmat(gv,1,mini); % top down sample of v gu = rand(ndimsq,mini) < sig(Gv + repmat(h,1,mini)); % top down sample of u rv = sig(Wgu + repmat(w,1,mini)); w = w + eta(v - rv)ones(mini,1); % delta rule for w W = W + eta(v - rv)gu'; % delta rule for W if (mod(it,200)==0) disp(sprintf('%d',it)); clf; colormap(gray); for i=1:nhid subplot(4,nhid,i); imagesc(reshape(G(:,i),ndim,ndim),[min(G(:)) max(G(:))]); set(gca,'xtick',[],'ytick',[]); title(num2str(sig(g(i)))); % title is gen prob colorbar; subplot(4,nhid,i+nhid); imagesc(reshape(W(i,:),ndim,ndim),[min(W(:)) max(W(:))]); set(gca,'xtick',[],'ytick',[]); colorbar; subplot(4,nhid,i+2nhid); imagesc(reshape(u(:,rp(i)),ndim,ndim),[0 1]); set(gca,'xtick',[],'ytick',[]); title(rp(i)); % title is pattern number colorbar; subplot(4,nhid,i+3nhid); imagesc(reshape(gpu(:,i),ndim,ndim),[0 1]); set(gca,'xtick',[],'ytick',[]); title(sprintf('%1d',sv(:,i))); % title is hidden unit state colorbar; end drawnow; endend
Expand Down
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		% c10p11.m%% demonstration of the Helmholtz machine using the wake-sleep% algorithm. rand('state',105);ndim=4; % dimension of the square of barsndimsq=ndim^2;nu=400; % number of examplespbar=0.5; % probability that a given bar is onpout=0.05; % extra noise added at the endii=ones(1,ndim);oo=zeros(1,ndim);conv=zeros(ndimsq,ndim);for i=1:ndim conv((i-1)ndim+(1:ndim),i)=ones(ndim,1);end bars=conv([rand(ndim/2,nu/2)<pbar zeros(ndim/2,nu/2); ... zeros(ndim/2,nu/2) rand(ndim/2,nu/2)<pbar]); % which bars are on in which examplesu=bars.(rand(ndimsq,nu)>pout) + (1-bars).(rand(ndimsq,nu)<pout); % generate output noise
Expand Down
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		% c10p6.m%% factor analysis exercisenu=500; % number of data pointsrandn('state',100); % to make the same as in the textgtrue=[1 ; 1 ; 1]; % since the underlying factor is identical in % each observable componentvtrue=randn(1,nu); u1=gtruevtrue + 0.5randn(3,nu); % data for 10.4A and Bu2=u1;u2(2,:)=vtrue+3randn(1,nu); % corrupt u_2 by more noiseC1=cov(u1'); % covariance matrix for u1C2=cov(u2'); % covariance matrix for u2 % the means are not very important since there % are so many data points.rot=[cos(pi/3) sin(pi/3) 0 ; -sin(pi/3) cos(pi/3) 0 ; 0 0 1] ... [cos(pi/6) 0 sin(pi/6) ; 0 1 0 ; -sin(pi/6) 0 cos(pi/6)];
Expand Down
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		% mapping from cortex into the retinalam=1.2 % constants for the mapep0=1.0;[X,Y]=meshgrid(0.05:.01:3,-2:.01:2); % global coordinates for the cortexep=ep0(exp(X/lam)-1); % eccentricitya=-(180(ep0+ep).Y)./(lameppi); % azimuthin=abs(a)<90; a=a.(abs(a)<90)+90(a>=90) - 90(a<=-90); % take care of absurditiesclf;K=8;Theta=pi/6;Phi=0;s=cos(KXcos(Theta)+KYsin(Theta)-Phi);subplot(2,2,1);imagesc(X(1,:),Y(:,1),s.in);set(gca,'fontsize',18,'fontname','palatino');title('visual cortex');xlabel('X (cm)');ylabel('Y (cm)');daspect([1 1.5 1]);subplot(2,2,2);contourf(ep.cos(a2pi/360),ep.sin(a2*pi/360),s,'w:');set(gca,'xlim',[0 12],'ylim',[-12 12]);set(gca,'fontsize',18,'fontname','palatino');title('visual space');daspect([1 1 1]);xlabel('degrees');ylabel('degrees');
Expand Down
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		function [est,K,ind]=c3p10(stim,spk,nfft)%[est,K,sstim]=wf(stim,spk,nfft)%Finds the optimal decoding filter K to predict stim from spk using %matlab function spectrum with nfft frequencies%est is the resulting stimulus estimate%K is the filter%ind is the indices such that est approximmates stim(ind).spk=spk-mean(spk); % remove the mean of the spikesmstim=mean(stim); stim=stim-mstim; % remove the mean of the stimulusP=spectrum(spk,stim,nfft); % work out the various self- and % cross-spectral termsTxy=P(:,4); % the complex transfer function from X to Y -- % ie the FFT of the decoding filterT=[Txy ; conj(Txy((end-1):-1:2))]; % want the real-valued versionK=fftshift(real(ifft(T)));% shift to make it centeredest=fftfilt(K,spk); % use the fast filter algorithmest=est(nfft+1:end)+mstim; % get rid of the transientind=(nfft/2+1):(length(stim)-nfft/2); % the indices of stim that % match those of est
Expand Down
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1 @@
		% c7p5.m % % symmetric recurrent nonlinear networkM=[0 -1 ; -1 0]; % connection matrixh=[1 ; 1]; % inputiters=100; % maximum number of iterationseta=0.1; % Euler stepclf;subplot(1,2,1);hold on;beta=2.0; % high gain for the activation function % now step the starting conditions around the outside % of a grid v1=0;for v2=0:.25:3 c7p5sub(M,beta,h,[v1 ; v2],eta,iters); drawnow;endv2=0;for v1=0:.25:3 c7p5sub(M,beta,h,[v1 ; v2],eta,iters); drawnow;endv1=3;for v2=0:.25:3 c7p5sub(M,beta,h,[v1 ; v2],eta,iters); drawnow;endv2=3;for v1=0:.25:3 c7p5sub(M,beta,h,[v1 ; v2],eta,iters); drawnow;end % now label the graphaxis('square'); grid;title(sprintf('beta=%f',beta),'FontSize',20);xlabel('v_1','FontSize',16); ylabel('v_2','FontSize',16);subplot(1,2,2);hold on;beta=0.5; % low gain for the activation function % now step the starting conditions around the outside % of a grid v1=0;for v2=0:.25:3 c7p5sub(M,beta,h,[v1 ; v2],eta,iters); drawnow;endv2=0;for v1=0:.25:3 c7p5sub(M,beta,h,[v1 ; v2],eta,iters); drawnow;endv1=3;for v2=0:.25:3 c7p5sub(M,beta,h,[v1 ; v2],eta,iters); drawnow;endv2=3;for v1=0:.25:3 c7p5sub(M,beta,h,[v1 ; v2],eta,iters); drawnow;endaxis('square'); grid;title(sprintf('beta=%f',beta),'FontSize',20);xlabel('v_1','FontSize',16); ylabel('v_2','FontSize',16);
Expand Down