how to predict from a trained neural network ?

To predict from a trained neural network in MATLAB, use classify(net, new_data) for classification tasks to get class labels, or predict(net, new_data) for regression tasks and raw probability scores. If using a shallow neural network (created with fitnet or patternnet), evaluate inputs directly by calling the network object like a function: output = net(input_matrix).

Core Methods Based on Network Type

1. Deep Classification Networks (DAGNetwork / SeriesNetwork): Use classify() to return categorical class predictions directly, or [labels, scores] = classify(net, X) to inspect the softmax probability distribution across classes.
2. Deep Regression Networks: Use YPred = predict(net, X) to generate continuous numerical estimates.
3. Shallow Neural Networks: Use Y = net(X) or Y = sim(net, X), where X has features along rows and observations along columns.

Executable MATLAB Code: End-to-End Prediction Pipeline

% =========================================================================
% METHOD 1: Predicting with Modern Deep Learning Networks (Images or Tables)
% =========================================================================

% 1. Load a pre-trained network from disk or workspace
loadedModel = load('trainedModel.mat'); 
net = loadedModel.net; % Contains trained DAGNetwork or SeriesNetwork

% 2. Read and prepare a new test sample
testImage = imread('new_sample.jpg');

% Match network input dimensions (e.g., 224x224 RGB)
inputSize = net.Layers(1).InputSize(1:2); 
preprocessedImage = imresize(testImage, inputSize);

% 3. Run inference
% Option A: For categorical class prediction
[predictedLabel, classProbabilities] = classify(net, preprocessedImage);

% Option B: For continuous numerical regression
% predictedValue = predict(net, preprocessedImage);

fprintf('Predicted Class: %s (Confidence: %.2f%%)\n', ...
    string(predictedLabel), max(classProbabilities) * 100);

% =========================================================================
% METHOD 2: Predicting with Shallow Neural Networks (feedforwardnet/patternnet)
% =========================================================================

% Suppose 'shallowNet' was trained with 5 input features
% Shape: [numFeatures x numSamples] -> Note: columns are observations
newSample = [12.4; 3.2; 0.85; 110.0; 4.5]; 

% If you applied normalization (e.g., mapminmax) during training, apply it here:
% newSampleNormalized = mapminmax('apply', newSample, trainingSettings);

% Run prediction
rawPrediction = net(newSample);

% For pattern recognition networks, find class with highest response
[~, classIndex] = max(rawPrediction);
fprintf('Shallow Network Predicted Class Index: %d\n', classIndex);

Crucial Engineering Checklist Before Running Predictions

• Input Dimension Alignment: Deep CNNs expect a 4D array for batches [Height, Width, Channels, BatchSize]. When passing a single image, MATLAB allows a 3D array [Height, Width, Channels], but tabular inputs to feature networks must strictly match the layer training dimensions.
• Consistent Preprocessing: If your network was trained on zero-mean normalized inputs or pixel values scaled to [0, 1], unscaled inputs (such as raw uint8 images ranging from 0 to 255) will produce completely incorrect predictions. Apply the exact same transformation function.
• GPU Acceleration for Batch Inference: When evaluating thousands of test records or video frames simultaneously, pass 'ExecutionEnvironment', 'gpu' into predict() to accelerate throughput via CUDA cores.