Fig. 1: Simplified model architecture diagrams.

The Black arrows represent the layers that connect each input to each output. The green arrows represent the positive feedback from backpropagation that aims to minimize error. The unfilled/white arrows represent stop gradient layers, which prevent the task from changing the weights in all layers upstream from the stop gradient layer. Red arrows represent gradient reversal layers of adversarial tasks, which reverse the direction of the weight changes and maximize loss for the task in any layer upstream from the gradient reversal layer. a represents the PC NN model, b represents the Geno-PC model, c represents the Geno-PC-T2D model, d represents the Geno-T2D Track-PC model, e represents the Geno-T2D Adv-PC model, f represents the CID-T2D Track-PC model, g represents the CID-T2D Track-PC/Geno model, h represents the CID Specific model, and i represents the Geno Specific model. An overview of these models can be found in Table 1. Note that some models, like (f) vs (g) and (h) vs (i), differ only in the way gradients flow during backpropagation, illustrated by the color of the arrows. The blue rectangles represent the input data, and the blue circles represent the transformed input that has been reduced to fewer dimensions. Task specific output layers are present for each task in each model. Principal component (PC), Type 2 Diabetes (T2D), Context Informed Data matrix (CID), Genotype (Geno), Adversarial (Adv), Convolutional (Conv).