Artificial intelligence justification

Machine learning methods

neuro uses a multilayer perceptron with two ReLU hidden layers. Parameters are trained on labeled examples using backpropagation, Adam and cross-entropy loss.

Recognition

The trained network classifies new sensor feature vectors. Decisions depend on learned weights, not only fixed thresholds. The output layer can adapt using external labeled feedback. The full justification is available as a separate Russian PDF.