We have just finished our first mile in the R and deep learning journey! Through this chapter, we got more familiar with the important concepts of deep learning. We started with what deep learning is all about, why it is important and the recent success of applications, as well. After we were well equipped, we solved the handwritten digit using shallow neural networks, deep neural networks and CNNs in sequence, and proved that CNNs are the best suited to exploiting strong and unique features that differentiate images of different classes.
Inspired by the human visual cortex, CNNs classify images by first deriving rich representations such as edges, curves and shapes, which was demonstrated in the visualization of the outputs of convolutional layers. In addition, we verified the performance and generalization of the CNN model using early stopping as a technique to avoid overfitting. Overall, we not only covered the mechanics of CNNs, including the concepts of convolution and pooling, but also implemented a CNN model with MXNet, as one of the most popular deep learning packages in R.
