Abstract:

We investigate the influence of batch normalization and spatial dropout on the performance of the VGG-11, Network in Network, and GoogLeNet architectures. We implement these models with the aforementioned regularization techniques and evaluate their impact on classification accuracy
and average loss. Our results indicate that spatial dropout has a limited effect on the evaluated metrics, whereas batch normalization yields a noticeable improvement in performance. Furthermore, we observe that combining both methods does not provide a significant advantage over the use of batch normalization alone.