AAI_2025_Capstone_Chronicles_Combined

To compare performance, Figure 5.5 contains the confusion matrices of the PyTorch model, the average confusion matrix of all the TurbaNet models in the swarm, and finally the confusion matrix of the ensemble model.

Figure 5.5 MNIST Confusion Matrices (left: PyTorch, middle: TurbaNet individual model average, right: TurbaNet Ensemble)

Figure 5.6 shows the differences taken between the confusion matrices to more clearly show the areas where the ensemble performs over the others. The left hand matrix is comparing the ensemble model to the average of the models. Ideally, the values on the diagonal should be positive and as high as possible and all terms off the diagonal should be negative and as low as possible. This corresponds to improving the number of correct predictions of the digits and reducing the incorrect predictions. As expected, the ensemble model outperforms the average of individual models—and even surpasses the baseline PyTorch model.

141