Multifidelity data fusion in convolutional encoder/decoder networks

We analyze the regression accuracy of convolutional neural networks assembled from encoders, decoders and skip connections trained with multifidelity data. Besides requiring significantly less trainable parameters than equivalent fully connected networks, encoder, decoder, encoder-decoder or decoder-encoder architectures can learn mapping between inputs to outputs arbitrary dimensionality. demonstrate their when on a few high-fidelity many low-fidelity data generated models ranging one-dimensional functions Poisson equation solvers in two-dimensions. finally discuss number implementation choices that improve reliability uncertainty estimates by Monte Carlo DropBlocks, compare among low-, high- approaches.