Figure 5
From: Enabling deeper learning on big data for materials informatics applications
Design approach for IRNet. Plain Network is composed of sequence of fully connected layer, where each layer is composed of a dense layer followed by batch normalization51 and ReLU54. Existing approach of residual learning places shortcut connection around each stack of multiple such layers where all the layers within each stack have same configuration (SRNet). The presented approach of individual residual network (IRNet) places shortcut connection around each layer which makes it easy for the model to learn the mapping of output materials property from the material composition and/or structure in the model input.
