Anatomy of CNNs
In the last section, we saw some of the challenges DNNs grappled with when dealing with visual recognition tasks. These issues include the lack of spatial awareness, high dimensionality, computational inefficiency, and the risk of overfitting. How do we overcome these challenges? This is where CNNs come into the picture. CNNs by design are uniquely positioned to handle image data. Let's go through Figure 7.1 and uncover why and how CNNs stand out:
Figure 7.1 – The anatomy of a CNN
Let’s break down the different layers in the diagram:
- Convolutional layer – the eyes of the network: Our journey begins with us feeding in images into the convolutional layer; this layer can be viewed as the “eyes of our network.” Their job is primarily to extract vital features. Unlike DNNs, where each neuron is connected to every neuron in the next layer, CNNs apply filters (also known as kernels) to capture local...
