Deep Learning Models: DenseNet and MobileNet

Your Name
Subject and Section
Professor’s Name
January 24, 2022
Neural Networks
DenseNet – Dense Convolutional Network (Image Classification)
Figure SEQ Figure \* ARABIC 1 - The Densenet Architecture | Image taken from Huang et al., 2017
DenseNet is a type of neural network that utilizes a series of blocks (i.e., Dense Blocks), made of relatively similar feature sizes, where our layers are connected to one another. One of the main reasons this type of neural network allows for strong foundational support for the development of AI technologies is that the series of layers receives “collective knowledge” from the others, known as the concatenation process. However, to fully understand the importance of Dense Net for developing any AI solution, a closer look at its features should be done.
One of the main elements of a Dense Net neural network is a Dense Block. A Dense Block refers to every one of the modules directly connecting all the layers in a traditional convolutional neural network. Figure 1 shows an image of a five-layer dense block with a growth rate of k = 4, which refers to the number of output maps of each layer within the network.
Additionally, it must be noted that each layer in this network should have similar feature sizes. The reason why layers are required to have similar or matching feature sizes is that the equation required for the process of concatenating or adding can only group the blocks if they are of the same size. In one study done by Li et al. (2019), the authors noted that feature sizes are crucial for preventing bottlenecks and disharmonies between variance shifts and maintaining the forward propagation of each Dense Net block. In other words, this supports the idea that the equation for the Dense Network would only work if there are similar feature sizes similar to what is presented in Figure 1.
MobileNet
Figure SEQ Figure \* ARABIC 2 - Mobile Net Model | Image is taken from Pujara, 2020
Mobile Net is another convolutional neural network used for AI and other machine learning processes. One of its main advantages is that it can process data, models, and systems at a higher speed and l...