How can we implement PCA for array of 512 feature maps of size 39*39 to get an output of array of x feature maps of size 39*39, where x is dimension desired after applying PCA.

How can we implement PCA for array of 512 feature maps of size 39*39 to get an output of array of x feature maps of size 39*39, where x is dimension desired after applying PCA. (Given x < 512)

Say I have a 3d matrix of size (39 39 512) and want to convert it into size (39 39 96) applying PCA. What is the matlab code for applying this without flattening the feature maps.

Answer 1

[Aditya]

Hi Shubham,

To apply PCA to a 3D matrix of feature maps without flattening them, you'll need to treat each 39x39 feature map as a single observation with 512 features. PCA will then reduce the dimensionality from 512 to the desired number of components (e.g., 96). Here's how you can implement this in MATLAB:

% Assume inputData is your 3D matrix of size (39, 39, 512)
inputData = rand(39, 39, 512);  % Example data, replace with your actual data

% Reshape the 3D matrix to a 2D matrix where each row is a feature map
[nRows, nCols, nFeatures] = size(inputData);
reshapedData = reshape(inputData, [], nFeatures);  % Resulting size will be (1521, 512)

% Apply PCA to reduce to 96 components
x = 96;  % Desired number of components
[coeff, score, ~, ~, explained] = pca(reshapedData);

% Take the first x principal components
reducedData = score(:, 1:x);  % Resulting size will be (1521, 96)

% Reshape back to 3D matrix of size (39, 39, 96)
outputData = reshape(reducedData, nRows, nCols, x);

% Display the percentage of variance explained by the selected components
disp(['Variance explained by the first ', num2str(x), ' components: ', num2str(sum(explained(1:x)))]);

% The outputData is now your reduced feature map array