This research project investigates the role of key factors that led to the resurgence of deep CNNs
and their success in classifying large datasets of natural images. Our investigation included the
role of new network components, the role of the training data, and the role of data augmentation.
Investigating the role of data augmentation led to the successful implementation of a deep CNN
that can be trained using a variable input size, which increased the amount of allowable scale
augmentation and led to much better single-view performance. Our analysis of the role of the
training data shows the capabilities of deep CNNs to break down a large hierarchical dataset
along the hierarchical lines into smaller components and learn all of them with great efficiency. This might help explain why deep CNN are very effective in classifying large and dense datasets of natural images which tend to have a hierarchical structure. Our investigation of core network components shows that the shared normalisation statistics of BN allowed us to alter the behaviour of the network by controlling the structure of the training batches. We used this observation to obtain large conditional gain by training and testing the network using balanced batches. Finally, we were able to implement a successful multitasking network that were able to outperform the corresponding single task networks. Our model used the normalisation statistics of BN to separate between the tasks, and our analysis shows that using a whole dataset per task increases the gains of the multitasking network by increasing the transfer of knowledge between the tasks.