Neural network (NN) model architectures used throughout this work. Rectangles correspond to regular dense layers with their corresponding number of nodes, elongated hexagons correspond to probabilistic dense variational layers, and circles correspond to skip connections. All hidden layers use ReLU activation functions except for those with an “x,” indicating no activation function is utilized. All models have five-band grizy magnitudes as inputs and output redshift predictions. We give more detail on the implementation in Section 3.2. Architecture (a) is for NN-1, (b) for NN-2 and NN-TL, (c) for NN-Combo, (d) for BNN-1, and (e) for BNN-2, BNN-TL, and BNN-Combo.