Analysing the empirical consequences of hidden layer configurations in ANNs

This paper investigates the impact of architectural variations in Artificial Neural Networks (ANNs), specifically focusing on the effects of hidden layer configurations-parallel, converging, and diverging, on classification performance. Each model was trained and evaluated under identical hyperparameter settings to ensure fair comparison. Evaluation metrics include precision, recall, F1-score, and a custom metric known as Signed Confidence Average (SCA), which incorporates prediction confidence into performance analysis. Experiments were conducted on the MNIST dataset to assess how different structural designs influence accuracy and confidence behavior in a controlled setting.