p300soft.zip_EEG预处理_matlab_脑电信号_脑电信号eeg_脑电信号处理

function b = train(b, x, y) % training procedure for Bayesian LDA % INPUT: % b - object of type bayeslda % x - m*n matrix containing n feature vectors of size m*1 % y - 1*n matrix containing class labels (-1,1) % % OUTPUT: % b - updated object of type bayeslda % % Author: Ulrich Hoffmann - EPFL, 2006 % Copyright: Ulrich Hoffmann - EPFL % % The algorithm implemented here was originally described by % MacKay, D. J. C., 1992. Bayesian interpolation. % Neural Computation 4 (3), pp. 415-447. %% compute regression targets from class labels (to do lda via regression) n_posexamples = sum(y==1); n_negexamples = sum(y==-1); n_examples = n_posexamples + n_negexamples; y(y==1) = n_examples/n_posexamples; y(y==-1) = -n_examples/n_negexamples; %% add feature that is constantly one (bias term) x = [x; ones(1,size(x,2))]; %% initialize variables for fast iterative estimation of alpha and beta n_features = size(x,1); % dimension of feature vectors d_beta = inf; % (initial) diff. between new and old beta d_alpha = inf; % (initial) diff. between new and old alpha alpha = 25; % (initial) inverse variance of prior distribution biasalpha = 0.00000001; % (initial) inverse variance of prior for bias term beta = 1; % (initial) inverse variance around targets stopeps = 0.0001; % desired precision for alpha and beta i = 1; % keeps track of number of iterations maxit = 500; % maximal number of iterations [v,d] = eig(x*x'); % needed for fast estimation of alpha and beta vxy = v'*x*y'; % dito d = diag(d); % dito e = ones(n_features-1,1); % dito %% estimate alpha and beta iteratively while ((d_alpha > stopeps) || (d_beta > stopeps)) && (i < maxit); alphaold = alpha; betaold = beta; m = beta*v*((beta*d+[alpha*e; biasalpha]).^(-1).*vxy); err = sum((y-m'*x).^2); gamma = sum(beta*d./(beta*d+[alpha*e; biasalpha])); alpha = gamma/(m'*m); beta = (n_examples - gamma)/err; if b.verbose fprintf('Iteration %i: alpha = %f, beta = %f\n',i,alpha,beta); end d_alpha = abs(alpha-alphaold); d_beta = abs(beta-betaold); i = i + 1; end %% process results of estimation if (i < maxit) % compute the log evidence % this can be used for simple model selection tasks % (see MacKays paper) b.evidence = (n_features/2)*log(alpha) + (n_examples/2)*log(beta) - ... (beta/2)*err - (alpha/2)*m'*m - ... 0.5*sum(log((beta*d+[alpha*e; biasalpha]))) - (n_examples/2)*log(2*pi); % store alpha, beta, the posterior mean and the posterrior precision- % matrix in class attributes b.alpha = alpha; b.beta = beta; b.w = m; b.p = v*diag((beta*d+[alpha*e; biasalpha]).^-1)*v'; if b.verbose fprintf('Optimization of alpha and beta successfull.\n'); fprintf('The logevidence is %f.\n',b.evidence); end else fprintf('Optimization of alpha and beta did not converge after %i iterations.\n',maxIt); fprintf('Giving up.'); end