-
Notifications
You must be signed in to change notification settings - Fork 0
/
spm_ADEM.m
874 lines (742 loc) · 31.4 KB
/
spm_ADEM.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
function [DEM] = spm_ADEM(DEM)
% Dynamic expectation maximisation: Active inversion
% FORMAT DEM = spm_ADEM(DEM)
%
% DEM.G - generative process
% DEM.M - recognition model
% DEM.C - causes
% DEM.U - prior expectation of causes
%__________________________________________________________________________
%
% This implementation of DEM is the same as spm_DEM but integrates both the
% generative process and model inversion in parallel. Its functionality is
% exactly the same apart from the fact that confounds are not accommodated
% explicitly. The generative model is specified by DEM.G and the veridical
% causes by DEM.C; these may or may not be used as priors on the causes for
% the inversion model DEM.M (i.e., DEM.U = DEM.C). Clearly, DEM.G does not
% require any priors or precision components; it will use the values of the
% parameters specified in the prior expectation fields.
%
% This routine is not used for model inversion per se but to simulate the
% dynamical inversion of models. Critically, it includes action
% variables a - that couple the model back to the generative process
% This enables active inference (c.f., action-perception) or embodied
% inference.
%
% hierarchical models M(i)
%--------------------------------------------------------------------------
% M(i).g = y(t) = g(x,v,P) {inline function, string or m-file}
% M(i).f = dx/dt = f(x,v,P) {inline function, string or m-file}
%
% M(i).pE = prior expectation of p model-parameters
% M(i).pC = prior covariances of p model-parameters
% M(i).hE = prior expectation of h hyper-parameters (cause noise)
% M(i).hC = prior covariances of h hyper-parameters (cause noise)
% M(i).gE = prior expectation of g hyper-parameters (state noise)
% M(i).gC = prior covariances of g hyper-parameters (state noise)
% M(i).Q = precision components (input noise)
% M(i).R = precision components (state noise)
% M(i).V = fixed precision (input noise)
% M(i).W = fixed precision (state noise)
% M(i).xP = precision (states)
%
% M(i).m = number of inputs v(i + 1);
% M(i).n = number of states x(i)
% M(i).l = number of output v(i)
% M(i).k = number of action a(i)
% hierarchical process G(i)
%--------------------------------------------------------------------------
% G(i).g = y(t) = g(x,v,a,P) {inline function, string or m-file}
% G(i).f = dx/dt = f(x,v,a,P) {inline function, string or m-file}
%
% G(i).pE = model-parameters
% G(i).U = precision (action)
% G(i).V = precision (input noise)
% G(i).W = precision (state noise)
%
% G(1).R = restriction or rate matrix for action [default: 1];
% G(i).aP = precision (action) [default: exp(-2)]
%
% G(i).m = number of inputs v(i + 1);
% G(i).n = number of states x(i)
% G(i).l = number of output v(i)
% G(i).k = number of action a(i)
%
%
% Returns the following fields of DEM
%--------------------------------------------------------------------------
%
% true model-states - u
%--------------------------------------------------------------------------
% pU.x = true hidden states
% pU.v = true causal states v{1} = response (Y)
% pU.C = prior covariance: cov(v)
% pU.S = prior covariance: cov(x)
%
% model-parameters - p
%--------------------------------------------------------------------------
% pP.P = parameters for each level
%
% hyper-parameters (log-transformed) - h,g
%--------------------------------------------------------------------------
% pH.h = cause noise
% pH.g = state noise
%
% conditional moments of model-states - q(u)
%--------------------------------------------------------------------------
% qU.a = Action
% qU.x = Conditional expectation of hidden states
% qU.v = Conditional expectation of causal states
% qU.z = Conditional prediction errors (v)
% qU.C = Conditional covariance: cov(v)
% qU.S = Conditional covariance: cov(x)
%
% conditional moments of model-parameters - q(p)
%--------------------------------------------------------------------------
% qP.P = Conditional expectation
% qP.C = Conditional covariance
%
% conditional moments of hyper-parameters (log-transformed) - q(h)
%--------------------------------------------------------------------------
% qH.h = Conditional expectation (cause noise)
% qH.g = Conditional expectation (state noise)
% qH.C = Conditional covariance
%
% F = log evidence = log marginal likelihood = negative free energy
%__________________________________________________________________________
%
% spm_ADEM implements a variational Bayes (VB) scheme under the Laplace
% approximation to the conditional densities of states (u), parameters (p)
% and hyperparameters (h) of any analytic nonlinear hierarchical dynamic
% model, with additive Gaussian innovations. It comprises three
% variational steps (D,E and M) that update the conditional moments of u, p
% and h respectively
%
% D: qu.u = max <L>q(p,h)
% E: qp.p = max <L>q(u,h)
% M: qh.h = max <L>q(u,p)
%
% where qu.u corresponds to the conditional expectation of hidden states x
% and causal states v and so on. L is the ln p(y,u,p,h|M) under the model
% M. The conditional covariances obtain analytically from the curvature of
% L with respect to u, p and h.
%
% The D-step is embedded in the E-step because q(u) changes with each
% sequential observation. The dynamical model is transformed into a static
% model using temporal derivatives at each time point. Continuity of the
% conditional trajectories q(u,t) is assured by a continuous ascent of F(t)
% in generalised co-ordinates. This means DEM can deconvolve online and
% represents an alternative to Kalman filtering or alternative Bayesian
% update procedures.
%
%__________________________________________________________________________
% Copyright (C) 2008 Wellcome Trust Centre for Neuroimaging
% Karl Friston
% $Id: spm_ADEM.m 7679 2019-10-24 15:54:07Z spm $
% check model, data, priors and unpack
%--------------------------------------------------------------------------
DEM = spm_ADEM_set(DEM);
M = DEM.M;
G = DEM.G;
C = DEM.C;
U = DEM.U;
% check whether to print
%--------------------------------------------------------------------------
try
db = DEM.db;
catch
db = 1;
end
% find or create a DEM figure
%--------------------------------------------------------------------------
if db
Fdem = spm_figure('GetWin','DEM');
end
% ensure embedding dimensions are compatible
%--------------------------------------------------------------------------
G(1).E.n = M(1).E.n;
G(1).E.d = M(1).E.n;
% order parameters (d = n = 1 for static models) and checks
%==========================================================================
d = M(1).E.d + 1; % embedding order of q(v)
n = M(1).E.n + 1; % embedding order of q(x)
s = M(1).E.s; % smoothness - s.d. (bins)
% number of states and parameters - generative model
%--------------------------------------------------------------------------
nY = size(C,2); % number of samples
nl = size(M,2); % number of levels
nv = sum(spm_vec(M.m)); % number of v (causal states)
nx = sum(spm_vec(M.n)); % number of x (hidden states)
ny = M(1).l; % number of y (inputs)
nc = M(end).l; % number of c (prior causes)
nu = nv*d + nx*n; % number of generalised states
% number of states and parameters - generative process
%--------------------------------------------------------------------------
gr = sum(spm_vec(G.l)); % number of v (outputs)
ga = sum(spm_vec(G.k)); % number of a (active states)
gx = sum(spm_vec(G.n)); % number of x (hidden states)
gy = G(1).l; % number of y (inputs)
na = ga; % number of a (action)
% number of iterations
%--------------------------------------------------------------------------
try, nE = M(1).E.nE; catch, nE = 16; end
try, nM = M(1).E.nM; catch, nM = 8; end
try, dt = M(1).E.dt; catch, dt = 1; end
% initialise regularisation parameters
%--------------------------------------------------------------------------
te = 2; % log integration time for E-Step
global t
% precision (roughness) of generalised fluctuations
%--------------------------------------------------------------------------
iV = spm_DEM_R(n,s);
iG = spm_DEM_R(n,s);
% time-delay operators (absorb motor delays into motor gain matrix)
%--------------------------------------------------------------------------
try
nG = norm(iG);
iG = iG*spm_DEM_T(n,-M(1).Ta);
iG = iG*nG/norm(iG);
end
try
Ty = spm_DEM_T(n,-M(1).Ty);
Ty = kron(Ty,speye(ny,ny));
end
% precision components Q{} requiring [Re]ML estimators (M-Step)
%==========================================================================
Q = {};
for i = 1:nl
q0{i,i} = sparse(M(i).l,M(i).l); %#ok<AGROW>
r0{i,i} = sparse(M(i).n,M(i).n);
end
Q0 = kron(iV,spm_cat(q0));
R0 = kron(iV,spm_cat(r0));
for i = 1:nl
for j = 1:length(M(i).Q)
q = q0;
q{i,i} = M(i).Q{j};
Q{end + 1} = blkdiag(kron(iV,spm_cat(q)),R0);
end
for j = 1:length(M(i).R)
q = r0;
q{i,i} = M(i).R{j};
Q{end + 1} = blkdiag(Q0,kron(iV,spm_cat(q)));
end
end
% and fixed components P
%--------------------------------------------------------------------------
Q0 = kron(iV,spm_cat(spm_diag({M.V})));
R0 = kron(iV,spm_cat(spm_diag({M.W})));
Qp = blkdiag(Q0,R0);
nh = length(Q); % number of hyperparameters
iR = [zeros(1,ny),ones(1,nv),ones(1,nx)]; % for empirical priors
iR = kron(speye(n,n),diag(iR));
% restriction or rate matrices - in terms of precision
%--------------------------------------------------------------------------
q0{1} = G(1).U;
Q0 = kron(iG,spm_cat(q0));
R0 = kron(iG,spm_cat(r0));
iG = blkdiag(Q0,R0);
% restriction or rate matrices - in terms of dE/da
%--------------------------------------------------------------------------
try
R = sparse(sum(spm_vec(G.l)),na);
R(1:ny,:) = G(1).R;
R = kron(spm_speye(n,1,0),R);
catch
R = 1;
end
% fixed priors on action (a)
%--------------------------------------------------------------------------
try
aP = G(1).aP;
catch
aP = exp(-2);
end
% fixed priors on states (u)
%--------------------------------------------------------------------------
xP = spm_cat(spm_diag({M.xP}));
Px = kron(iV(1:n,1:n),speye(nx,nx)*exp(-8) + xP);
Pv = kron(iV(1:d,1:d),speye(nv,nv)*exp(-8));
Pa = spm_speye(na,na)*aP;
Pu = spm_cat(spm_diag({Px Pv}));
% hyperpriors
%--------------------------------------------------------------------------
ph.h = spm_vec({M.hE M.gE}); % prior expectation of h
ph.c = spm_cat(spm_diag({M.hC M.gC})); % prior covariances of h
qh.h = ph.h; % conditional expectation
qh.c = ph.c; % conditional covariance
ph.ic = spm_inv(ph.c); % prior precision
% priors on parameters (in reduced parameter space)
%==========================================================================
pp.c = cell(nl,nl);
qp.p = cell(nl,1);
for i = 1:(nl - 1)
% eigenvector reduction: p <- pE + qp.u*qp.p
%----------------------------------------------------------------------
qp.u{i} = spm_svd(M(i).pC); % basis for parameters
M(i).p = size(qp.u{i},2); % number of qp.p
qp.p{i} = sparse(M(i).p,1); % initial qp.p
pp.c{i,i} = qp.u{i}'*M(i).pC*qp.u{i}; % prior covariance
try
qp.e{i} = qp.p{i} + qp.u{i}'*(spm_vec(M(i).P) - spm_vec(M(i).pE));
catch
qp.e{i} = qp.p{i}; % initial qp.e
end
end
Up = spm_cat(spm_diag(qp.u));
% initialise and augment with confound parameters B; with flat priors
%--------------------------------------------------------------------------
np = sum(spm_vec(M.p)); % number of model parameters
pp.c = spm_cat(pp.c);
pp.ic = spm_inv(pp.c);
% initialise conditional density q(p) (for D-Step)
%--------------------------------------------------------------------------
qp.e = spm_vec(qp.e);
qp.c = sparse(np,np);
% initialise cell arrays for D-Step; e{i + 1} = (d/dt)^i[e] = e[i]
%==========================================================================
qu.x = cell(n,1);
qu.v = cell(n,1);
qu.a = cell(1,1);
qu.y = cell(n,1);
qu.u = cell(n,1);
pu.v = cell(n,1);
pu.x = cell(n,1);
pu.z = cell(n,1);
pu.w = cell(n,1);
[qu.x{:}] = deal(sparse(nx,1));
[qu.v{:}] = deal(sparse(nv,1));
[qu.a{:}] = deal(sparse(na,1));
[qu.y{:}] = deal(sparse(ny,1));
[qu.u{:}] = deal(sparse(nc,1));
[pu.v{:}] = deal(sparse(gr,1));
[pu.x{:}] = deal(sparse(gx,1));
[pu.z{:}] = deal(sparse(gr,1));
[pu.w{:}] = deal(sparse(gx,1));
% initialise cell arrays for hierarchical structure of x[0] and v[0]
%--------------------------------------------------------------------------
qu.x{1} = spm_vec({M(1:end - 1).x});
qu.v{1} = spm_vec({M(1 + 1:end).v});
qu.a{1} = spm_vec({G.a});
pu.x{1} = spm_vec({G.x});
pu.v{1} = spm_vec({G.v});
% derivatives for Jacobian of D-step
%--------------------------------------------------------------------------
Dx = kron(spm_speye(n,n,1),spm_speye(nx,nx,0));
Dv = kron(spm_speye(d,d,1),spm_speye(nv,nv,0));
Dc = kron(spm_speye(d,d,1),spm_speye(nc,nc,0));
Da = kron(spm_speye(1,1,1),sparse(na,na));
Du = spm_cat(spm_diag({Dx,Dv}));
Dq = spm_cat(spm_diag({Dx,Dv,Dc,Da}));
Dx = kron(spm_speye(n,n,1),spm_speye(gx,gx,0));
Dv = kron(spm_speye(n,n,1),spm_speye(gr,gr,0));
Dp = spm_cat(spm_diag({Dv,Dx,Dv,Dx}));
dfdw = kron(speye(n,n),speye(gx,gx));
dydv = kron(speye(n,n),speye(gy,gr));
% and null blocks
%--------------------------------------------------------------------------
dVdc = sparse(d*nc,1);
% gradients and curvatures for conditional uncertainty
%--------------------------------------------------------------------------
dWdu = sparse(nu,1);
dWduu = sparse(nu,nu);
% preclude unnecessary iterations
%--------------------------------------------------------------------------
if ~np && ~nh, nE = 1; end
% create innovations (and add causes)
%--------------------------------------------------------------------------
[z,w] = spm_DEM_z(G,nY);
z{end} = C + z{end};
a = {G.a};
Z = spm_cat(z(:));
W = spm_cat(w(:));
A = spm_cat(a(:));
% Iterate DEM
%==========================================================================
F = -Inf;
for iE = 1:nE
% get time and clear persistent variables in evaluation routines
%----------------------------------------------------------------------
tic; clear spm_DEM_eval
% E-Step: (with embedded D-Step)
%======================================================================
% [re-]set accumulators for E-Step
%----------------------------------------------------------------------
dFdp = zeros(np,1);
dFdpp = zeros(np,np);
EE = sparse(0);
ECE = sparse(0);
EiSE = sparse(0);
qp.ic = sparse(0);
Hqu.c = sparse(0);
% [re-]set precisions using [hyper]parameter estimates
%----------------------------------------------------------------------
iS = Qp;
for i = 1:nh
iS = iS + Q{i}*exp(qh.h(i));
end
% precision for empirical priors
%----------------------------------------------------------------------
iP = iR*iS*iR;
% [re-]set states & their derivatives
%----------------------------------------------------------------------
try
qu = qU(1);
pu = pU(1);
end
% D-Step: (nY samples)
%======================================================================
for iY = 1:nY
% time (GLOBAL variable for non-automomous systems)
%------------------------------------------------------------------
t = iY/nY;
% pass action to pu.a (external states)
%==================================================================
try, A = spm_cat({qU.a qu.a}); end
% derivatives of responses and random fluctuations
%------------------------------------------------------------------
pu.z = spm_DEM_embed(Z,n,iY);
pu.w = spm_DEM_embed(W,n,iY);
pu.a = spm_DEM_embed(A,n,iY);
qu.u = spm_DEM_embed(U,n,iY);
% evaluate generative process
%------------------------------------------------------------------
[pu,dg,df] = spm_ADEM_diff(G,pu);
% and pass response to qu.y
%==================================================================
for i = 1:n
y = spm_unvec(pu.v{i},{G.v});
qu.y{i} = y{1};
end
% sensory delays
%------------------------------------------------------------------
try, qu.y = spm_unvec(Ty*spm_vec(qu.y),qu.y); end
% evaluate generative model
%------------------------------------------------------------------
[E,dE] = spm_DEM_eval(M,qu,qp);
% conditional covariance [of states {u}]
%------------------------------------------------------------------
qu.c = spm_inv(dE.du'*iS*dE.du + Pu);
pu.c = spm_inv(dE.du'*iP*dE.du + Pu);
Hqu.c = Hqu.c + spm_logdet(qu.c);
% save at qu(t)
%------------------------------------------------------------------
qE{iY} = E;
qC{iY} = qu.c;
pC{iY} = pu.c;
qU(iY) = qu;
pU(iY) = pu;
% and conditional precision
%------------------------------------------------------------------
if nh
ECEu = dE.du*qu.c*dE.du';
ECEp = dE.dp*qp.c*dE.dp';
end
% uncertainty about parameters dWdv, ... ; W = ln(|qp.c|)
%==================================================================
if np
for i = 1:nu
CJp(:,i) = spm_vec(qp.c*dE.dpu{i}'*iS);
dEdpu(:,i) = spm_vec(dE.dpu{i}');
end
dWdu = CJp'*spm_vec(dE.dp');
dWduu = CJp'*dEdpu;
end
% tensor products for Jacobian (generative process)
%------------------------------------------------------------------
Dgda = kron(spm_speye(n,1,1),dg.da);
Dgdv = kron(spm_speye(n,n,1),dg.dv);
Dgdx = kron(spm_speye(n,n,1),dg.dx);
dfda = kron(spm_speye(n,1,0),df.da);
dfdv = kron(spm_speye(n,n,0),df.dv);
dfdx = kron(spm_speye(n,n,0),df.dx);
dgda = kron(spm_speye(n,1,0),dg.da);
dgdx = kron(spm_speye(n,n,0),dg.dx);
% change in error w.r.t. action
%------------------------------------------------------------------
Dfdx = 0;
for i = 1:n
Dfdx = Dfdx + kron(spm_speye(n,n,-i),df.dx^(i - 1));
end
% dE/da with restriction (R)
%------------------------------------------------------------------
dE.dv = dE.dy*dydv;
dE.da = dE.dv*((dgda + dgdx*Dfdx*dfda).*R);
% first-order derivatives
%------------------------------------------------------------------
dVdu = -dE.du'*iS*E - Pu*spm_vec({qu.x{1:n} qu.v{1:d}}) - dWdu/2;
dVda = -dE.da'*iG*E - Pa*spm_vec( qu.a{1:1});
% and second-order derivatives
%------------------------------------------------------------------
dVduu = -dE.du'*iS*dE.du - Pu - dWduu/2 ;
dVdaa = -dE.da'*iG*dE.da - Pa;
dVduv = -dE.du'*iS*dE.dv;
dVduc = -dE.du'*iS*dE.dc;
dVdua = -dE.du'*iS*dE.da;
dVdav = -dE.da'*iG*dE.dv;
dVdau = -dE.da'*iG*dE.du;
dVdac = -dE.da'*iG*dE.dc;
% D-step update: of causes v{i}, and hidden states x(i)
%==================================================================
% states and conditional modes
%------------------------------------------------------------------
p = {pu.v{1:n} pu.x{1:n} pu.z{1:n} pu.w{1:n}};
q = {qu.x{1:n} qu.v{1:d} qu.u{1:d} qu.a{1:1}};
u = [p q];
% gradient
%------------------------------------------------------------------
dFdu = [ Dp*spm_vec(p);
spm_vec({dVdu; dVdc; dVda}) + Dq*spm_vec(q)];
% Jacobian (variational flow)
%------------------------------------------------------------------
dFduu = spm_cat(...
{Dgdv Dgdx Dv [] [] [] Dgda;
dfdv dfdx [] dfdw [] [] dfda;
[] [] Dv [] [] [] [];
[] [] [] Dx [] [] [];
dVduv [] [] [] Du+dVduu dVduc dVdua;
[] [] [] [] [] Dc []
dVdav [] [] [] dVdau dVdac dVdaa});
% update states q = {x,v,z,w} and conditional modes
%==================================================================
du = spm_dx(dFduu,dFdu,dt);
u = spm_unvec(spm_vec(u) + du,u);
% and save them
%------------------------------------------------------------------
pu.v(1:n) = u((1:n));
pu.x(1:n) = u((1:n) + n);
qu.x(1:n) = u((1:n) + n + n + n + n);
qu.v(1:d) = u((1:d) + n + n + n + n + n);
qu.a(1:1) = u((1:1) + n + n + n + n + n + d + d);
% Gradients and curvatures for E-Step: W = tr(C*J'*iS*J)
%==================================================================
if np
for i = 1:np
CJu(:,i) = spm_vec(qu.c*dE.dup{i}'*iS);
dEdup(:,i) = spm_vec(dE.dup{i}');
end
dWdp = CJu'*spm_vec(dE.du');
dWdpp = CJu'*dEdup;
% Accumulate; dF/dP = <dL/dp>, dF/dpp = ...
%--------------------------------------------------------------
dFdp = dFdp - dWdp/2 - dE.dp'*iS*E;
dFdpp = dFdpp - dWdpp/2 - dE.dp'*iS*dE.dp;
qp.ic = qp.ic + dE.dp'*iS*dE.dp;
end
% accumulate SSE
%------------------------------------------------------------------
EiSE = EiSE + E'*iS*E;
% and quantities for M-Step
%------------------------------------------------------------------
if nh
EE = E*E'+ EE;
ECE = ECE + ECEu + ECEp;
end
if nE == 1
% evaluate objective function (F)
%======================================================================
J(iY) = - trace(E'*iS*E)/2 ... % states (u)
+ spm_logdet(qu.c) ... % entropy q(u)
+ spm_logdet(iS)/2; % entropy - error
end
end % sequence (nY)
% augment with priors
%----------------------------------------------------------------------
dFdp = dFdp - pp.ic*qp.e;
dFdpp = dFdpp - pp.ic;
qp.ic = qp.ic + pp.ic;
qp.c = spm_inv(qp.ic);
% E-step: update expectation (p)
%======================================================================
% update conditional expectation
%----------------------------------------------------------------------
dp = spm_dx(dFdpp,dFdp,{te});
qp.e = qp.e + dp;
qp.p = spm_unvec(qp.e,qp.p);
% M-step - hyperparameters (h = exp(l))
%======================================================================
mh = zeros(nh,1);
dFdh = zeros(nh,1);
dFdhh = zeros(nh,nh);
for iM = 1:nM
% [re-]set precisions using [hyper]parameter estimates
%------------------------------------------------------------------
iS = Qp;
for i = 1:nh
iS = iS + Q{i}*exp(qh.h(i));
end
S = spm_inv(iS);
dS = ECE + EE - S*nY;
% 1st-order derivatives: dFdh = dF/dh
%------------------------------------------------------------------
for i = 1:nh
dPdh{i} = Q{i}*exp(qh.h(i));
dFdh(i,1) = -trace(dPdh{i}*dS)/2;
end
% 2nd-order derivatives: dFdhh
%------------------------------------------------------------------
for i = 1:nh
for j = 1:nh
dFdhh(i,j) = -trace(dPdh{i}*S*dPdh{j}*S*nY)/2;
end
end
% hyperpriors
%------------------------------------------------------------------
qh.e = qh.h - ph.h;
dFdh = dFdh - ph.ic*qh.e;
dFdhh = dFdhh - ph.ic;
% update ReML estimate of parameters
%------------------------------------------------------------------
dh = spm_dx(dFdhh,dFdh);
qh.h = qh.h + dh;
mh = mh + dh;
% conditional covariance of hyperparameters
%------------------------------------------------------------------
qh.c = -spm_inv(dFdhh);
% convergence (M-Step)
%------------------------------------------------------------------
if (dFdh'*dh < 1e-2) || (norm(dh,1) < exp(-8)), break, end
end % M-Step
% evaluate objective function (F)
%======================================================================
L = - trace(EiSE)/2 ... % states (u)
- trace(qp.e'*pp.ic*qp.e)/2 ... % parameters (p)
- trace(qh.e'*ph.ic*qh.e)/2 ... % hyperparameters (h)
+ Hqu.c/2 ... % entropy q(u)
+ spm_logdet(qp.c)/2 ... % entropy q(p)
+ spm_logdet(qh.c)/2 ... % entropy q(h)
- spm_logdet(pp.c)/2 ... % entropy - prior p
- spm_logdet(ph.c)/2 ... % entropy - prior h
+ spm_logdet(iS)*nY/2 ... % entropy - error
- n*ny*nY*log(2*pi)/2;
% if F is increasing, save expansion point and derivatives
%----------------------------------------------------------------------
if L > F(end) || iE < 3
% save model-states (for each time point)
%==================================================================
for t = 1:length(qU)
% states
%--------------------------------------------------------------
a = spm_unvec(qU(t).a{1},{G.a});
v = spm_unvec(pU(t).v{1},{G.v});
x = spm_unvec(pU(t).x{1},{G.x});
z = spm_unvec(pU(t).z{1},{G.v});
w = spm_unvec(pU(t).w{1},{G.x});
for i = 1:nl
try
PU.v{i}(:,t) = spm_vec(v{i});
PU.z{i}(:,t) = spm_vec(z{i});
end
try
PU.x{i}(:,t) = spm_vec(x{i});
PU.w{i}(:,t) = spm_vec(w{i});
end
try
QU.a{i}(:,t) = spm_vec(a{i});
end
end
% conditional modes
%--------------------------------------------------------------
v = spm_unvec(qU(t).v{1},{M(1 + 1:end).v});
x = spm_unvec(qU(t).x{1},{M(1:end - 1).x});
z = spm_unvec(qE{t}(1:(ny + nv)),{M.v});
w = spm_unvec(qE{t}((1:nx) + (ny + nv)*n),{M.x});
for i = 1:(nl - 1)
if M(i).m, QU.v{i + 1}(:,t) = spm_vec(v{i}); end
if M(i).l, QU.z{i}(:,t) = spm_vec(z{i}); end
if M(i).n, QU.x{i}(:,t) = spm_vec(x{i}); end
if M(i).n, QU.w{i}(:,t) = spm_vec(w{i}); end
end
QU.v{1}(:,t) = spm_vec(qU(t).y{1}) - spm_vec(z{1});
QU.z{nl}(:,t) = spm_vec(z{nl});
% and conditional covariances
%--------------------------------------------------------------
i = (1:nx);
QU.S{t} = qC{t}(i,i);
PU.S{t} = pC{t}(i,i);
i = (1:nv) + nx*n;
QU.C{t} = qC{t}(i,i);
PU.C{t} = pC{t}(i,i);
end
% save conditional densities
%------------------------------------------------------------------
B.QU = QU;
B.PU = PU;
B.qp = qp;
B.qh = qh;
% decrease regularisation
%------------------------------------------------------------------
F(iE) = L;
te = min(te + 1,8);
else
% otherwise, return to previous expansion point and break
%------------------------------------------------------------------
QU = B.QU;
PU = B.PU;
qp = B.qp;
qh = B.qh;
% increase regularisation
%------------------------------------------------------------------
F(iE) = F(end);
te = min(te - 1,0);
end
% report and break if convergence
%======================================================================
if db
figure(Fdem)
spm_DEM_qU(QU)
if np
subplot(nl,4,4*nl)
bar(full(Up*qp.e))
xlabel({'parameters';'{minus prior}'})
axis square, grid on
end
if length(F) > 2
subplot(nl,4,4*nl - 1)
plot(F - F(1))
xlabel('updates')
title('log-evidence')
axis square, grid on
end
drawnow
% report (EM-Steps)
%------------------------------------------------------------------
str{1} = sprintf('ADEM: %i (%i)',iE,iM);
str{2} = sprintf('F:%.4e',full(L - F(1)));
str{3} = sprintf('p:%.2e',full(dp'*dp));
str{4} = sprintf('h:%.2e',full(mh'*mh));
str{5} = sprintf('(%.2e sec)',full(toc));
fprintf('%-16s%-16s%-14s%-14s%-16s\n',str{:})
end
if (norm(dp,1) < exp(-8)) && (norm(mh,1) < exp(-8)), break, end
end
% assemble output arguments
%==========================================================================
% conditional moments of model-parameters (rotated into original space)
%--------------------------------------------------------------------------
qP.P = spm_unvec(Up*qp.e + spm_vec(M.pE),M.pE);
qP.C = Up*qp.c*Up';
qP.V = spm_unvec(diag(qP.C),M.pE);
% conditional moments of hyper-parameters (log-transformed)
%--------------------------------------------------------------------------
qH.h = spm_unvec(qh.h,{{M.hE} {M.gE}});
qH.g = qH.h{2};
qH.h = qH.h{1};
qH.C = qh.c;
qH.V = spm_unvec(diag(qH.C),{{M.hE} {M.gE}});
qH.W = qH.V{2};
qH.V = qH.V{1};
% Fill in DEM with response and its causes
%--------------------------------------------------------------------------
DEM.pP.P = {G.pE}; % parameters encoding process
DEM.M = M; % generative model
DEM.U = U; % causes
DEM.Y = PU.v{1}; % response
DEM.pU = PU; % prior moments of model-states
DEM.qU = QU; % conditional moments of model-states
DEM.qP = qP; % conditional moments of model-parameters
DEM.qH = qH; % conditional moments of hyper-parameters
DEM.F = F; % [-ve] Free energy
try
DEM.J = J; % [-ve] Free energy (over samples)
end