collapse across dimensions of a matrix

I have a multidimensional matrix 19 x 64 x 5 x 21 which corresponds to 19 participants x 64 channels x 5 timepoints x 21 data values.

I want to collapse the last two dimensions of the matrix so that I end up with the mean of the (5 x 21) data values, per participant, per channel (i.e. a 19 x 64 (x1) matrix)

I've tried using mean, tried using reshape etc but I am really struggling with the scale of the data and picturing how the the four dimensions fit together. As such I can't confidently manipulate the data across the four dimensions so any guidance would be appreciated.

Answer 1

[Steven Lord]

If you're using release R2018b or later, several data analysis functions (including mean) can accept a vector of dimensions over which to operate. See the Release Notes for a list of which functions can accept a vector of dimensions.

[madhan ravi]

+1

mean(rand(19 , 64 , 5 , 21),[3 4])  % last two dimensions

[MsPop]

this would be perfect but doesn't appear to be implemented yet for nanmean?

[Steven Lord]

Using madhan ravi's example, slightly modified:

First generate a sample array.

R = rand(19, 64, 5, 21);

Next add in a couple NaN values at random locations. I made a copy of R with which to work so you can compare them later.

nanloc = randperm(numel(R), 30);
R2 = R;
R2(nanloc) = NaN;

Finally take the mean ignoring NaN values.

result = mean(R2, [3 4], 'omitnan');

[MsPop]

Thank you so much - works perfectly and so very simple :) !!

Answer 2

[Mark Sherstan]

Give this a try:

% Generate random matrix
A = randi(10,[19 64]);
B = cat(3,A,A,A,A);
C = cat(4,B,B);

% Get indices and set counter
[idx1 idx2 idx3 idx4] = size(C);
counter = 1;

% Build cell array of all the matrices
for ii = 1:idx4
    for jj = 1:idx3
        out{counter} = C(:,:,idx3,idx4);
        counter = counter + 1;
    end
end

% Calculate the mean 
dim = ndims(out{1});
M = cat(dim+1,out{:});
meanMatrix = mean(M,dim+1);

[MsPop]

Many thanks, that appears to have worked!

Update: I'm manually calculating the mean value for one participant at one channel to double check that this is working (i.e. the mean across the 105 values) but it unfortunately doesn't tally. I think that it goes wrong in the build of the out variable, because the resulting 19x64 array is repeated across every one of the 105 cells. I reran the original code to check but it is hard to check there because of the way the matrices have been built. But I think that the following replicates the problem:

%% Generate random matrix

A = randi(10,[19 64]);

B = randi(10,[19 64]);

BB = cat(3,A,B,A,B);

CC = cat(3,B,A,B,A);

C = cat(4,BB,CC);

% Get indices and set counter

[idx1 idx2 idx3 idx4] = size(C);

counter = 1;

% Build cell array of all the matrices

for ii = 1:idx4

for jj = 1:idx3

out{counter} = C(:,:,idx3,idx4);

counter = counter + 1;

end

end

% Calculate the mean

dim = ndims(out{1});

M = cat(dim+1,out{:});

meanMatrix = mean(M,dim+1);

Answer 3

[madhan ravi]

doc permute

Answer 4

[Matt J]

mean(yourMatrix(:,:,:),3,'omitnan')

Answer 5

[Matt J]

Using sepblockfun (Download),

sepblockfun(yourArray, [1,1,inf,inf], 'mean')