Abstract
Accuracy and loss summarize predictive performance but provide limited information about how individual components evolve during training. This study presents a non-intrusive observation framework that records epoch-level descriptors of neuron activations, connection weights, gradients, updates, and biases without changing the training objective or optimizer. The framework was evaluated in five controlled experiments on multilayer perceptrons trained on MNIST and the Wisconsin Diagnostic Breast Cancer dataset. The final protocol comprised five random seeds, 30 MNIST epochs, 100 WDBC epochs, and 130 dataset-condition-seed combinations examining repeated runs, network width, learning rate, activation function, and descriptor-guided pruning. Models with similar validation accuracy frequently occupied different internal regimes. On MNIST, the coefficient of variation of gradient magnitude across seeds was 0.239, compared with 0.0021 for accuracy. Increasing learning rate from 0.0001 to 0.01 increased persistent inactivity from 0.0052 to 0.3656, while four activation functions separated by less than one percentage point of accuracy exhibited distinct inactivity and saturation trajectories. In the functional evaluation, descriptor-guided edge pruning caused less accuracy loss than layer-matched random pruning in all 15 MNIST seed-fraction comparisons and in 9 of 15 WDBC comparisons. However, the composite score and weight-magnitude pruning performed comparably at 5% to 20% sparsity. The findings show that component-level temporal descriptors reveal localized and reversible processes that are obscured by final global metrics. They support the use of internal dynamics as a complementary diagnostic layer, but do not establish a universal predictor of model quality or a pruning criterion superior to magnitude.