Augmentations: An Insight into Their Effectiveness on Convolution Neural Networks

Augmentations are the key factor in determining performance of any neural network as they provide a model with critical edge boosting its performance. Their ability to boost model’s robustness depends on two factors, viz-a-viz, architecture, and type augmentations. very specific dataset, it is not imperative that all kinds augmentation would necessarily produce positive effect Hence there need identify augmentations perform consistently well across variety datasets also remain invariant convolutions, number parameters used. This paper evaluates using 3 × depth-wise separable convolutions different techniques MNIST, FMNIST, CIFAR10 datasets. Statistical Evidence shows such Cutouts Random horizontal flip were consistent both parametrically low high architectures. Depth-wise outperformed at higher due their create deeper networks. resulted bridging accuracy gap between thus establishing role generalization. At did significant change The synergistic multiple parameters, antagonistic lower was evaluated. work proves delicate balance architectural supremacy needs be achieved enhance given deep learning task.