Single-Task Multi-Decoder Network Ensembles

Ensemble methods show superior performance, uncertainty estimates, and out-of-distribution robustness. At the same time, they have both a higher memory footprint and latency when it comes to deployment. To overcome the latter while still benefit from the former, we investigated in this project the use of single-task multi-decoder networks. Here, we consider the decoders as the ensemble and use a shared encoder instead. To this end, we investigate variations of the loss, architecture, and data augmentation.

This project was a joint effort with CARIAD SE.