Multi-branch Classifier (MC)
Inspired by Inception Module and multi-models, MC takes multi-branch structure to provide different views and enrich representation as well. In detail, MC consists of four FC layers.
In our study, MC only does linear projection at each branch instead of convolutional operation or the neural network, although it can be Conv \(1\times 1\).
The dimension change is 64 \(\rightarrow\) 16 \(\rightarrow\) [8,8,8,8] \(\rightarrow\) concatenate to 32 \(\rightarrow\) num\_class, where num\_class represents the number of class. In our experiment, it is 2. When we convert the dimension from 16 to 32, the above figure shows that we apply multi-branch structure to convert the dimension to 8 and repeat 4 times. Then, we concantenate them as a 32-dimensional tensor.
The multi-branch structure broadens the network. MC can provide more features and view the object from different angles.