Figure 6
Summarizes the concept of the PlexusNet architecture. A block consists of multiple neural network layers. Four architecture block types are available: VGG, Inception, residual, or soft attention block. The major hyperparameters for the graph definition are the depth (number of levels), with a minimum depth of 2; the number of end-to-end paths (width); the number of transitory “short” paths; and junctions that intersect between two end-to-end paths. Here, the example PlexusNet architecture has a depth of 3 levels, a width of 2, and a single transitory path and weighted junction between two end-to-end paths. The position of the weighted junction between two paths before the global pooling layer is determined randomly. For all the PlexusNet models, all the final feature maps of the end-to-end paths are concatenated before being fed into the global pooling layer. The depth of a transitory path is determined randomly, and the transitory path concatenates with the root path (by default, the first end-to-end path is considered the root path) at the same level as the weighted junction. The position randomization for weighted junctions or the depth for transitory paths has no impact on the model performance, while the number of weighted junctions has an impact on the model performance. For simplified model development, we unified the block type for all paths.
