Application of Basic Neural Computation Models: Perceptron and ADALINE

This work involves the implementation, training, and evaluation of two fundamental neural network models: the Perceptron and the Adaptive Linear Neural Network (ADALINE). The objective was to analyze the performance of these models in pattern recognition and time series forecasting tasks. The Perceptron used a single-layer architecture to classify numerical patterns, with its convergence and robustness to noise evaluated using different learning rates. ADALINE, a linear model, was implemented to predict values in a time series, similarly analyzing the impact of learning rate and noise. Results indicate that the Perceptron perfectly classifies linearly separable patterns without noise, but its performance severely degrades (accuracy ≈19%) with 20% noise. ADALINE also achieves a low forecasting error on clean data but demonstrates a high sensitivity to noise. It is concluded that while both models are effective for idealized and linearly separable problems, their capacity for generalization and robustness is limited, highlighting the necessity for more complex architectures for real-world applications.