Fixing "Index Exceeds Matrix Dimensions" in MATLAB Neural Networks

This error occurs in MATLAB Neural Network Toolbox (Deep Learning Toolbox) when data orientations, sample counts, or target formats do not match the dimensions expected by the network.

1. Cause #1: Samples vs. Features Orientation (Most Common)

MATLAB standard neural networks (fitnet, patternnet, feedforwardnet) require features in rows and samples in columns.

% Suppose X has 500 samples and 4 features (500x4)
% Suppose Y has 500 target values (500x1)

% Transpose to [Features x Samples]
X_train = X'; % Now 4 x 500
Y_train = Y'; % Now 1 x 500
net = fitnet(10);
net = train(net, X_train, Y_train);

2. Cause #2: Incorrect Indexing of Training Record (tr)

When you extract partition indices (tr.trainInd, tr.valInd, tr.testInd), remember that samples are stored in columns.

% Correct: Index columns
train_X = X_train(:, tr.trainInd);
train_Y = Y_train(:, tr.trainInd);

% Wrong: Indexing rows causes "Index exceeds matrix dimensions"
% train_X = X_train(tr.trainInd, :);

3. Cause #3: Target Label Format in Classification (patternnet)

For multi-class classification with \(C\) classes, patternnet requires a one-hot encoded matrix of size [C x N_samples], not a single vector of integer class labels (1, 2, 3).

% If your labels are class indices: labels = [1; 2; 3; 1; 2; ...]; (1 x N)
% Convert to one-hot binary vectors using ind2vec:
Y_onehot = full(ind2vec(labels)); % Produces [NumClasses x NumSamples]

net = patternnet(10);
net = train(net, X_train, Y_onehot);

4. Cause #4: Sample Count Mismatch Between X and Y

The number of columns in X and Y must be identical.

% Check dimensions before training
fprintf('X size: %d features x %d samples\n', size(X_train, 1), size(X_train, 2));
fprintf('Y size: %d outputs  x %d samples\n', size(Y_train, 1), size(Y_train, 2));

if size(X_train, 2) ~= size(Y_train, 2)
    error('Number of samples in X and Y must match.');
end

Quick Dimension Reference Table