Packed-Ensembles for Efficient Uncertainty Estimation

  • 2023-03-30 18:50:19
  • Olivier Laurent, Adrien Lafage, Enzo Tartaglione, Geoffrey Daniel, Jean-Marc Martinez, Andrei Bursuc, Gianni Franchi
  • 0

Abstract

Deep Ensembles (DE) are a prominent approach for achieving excellentperformance on key metrics such as accuracy, calibration, uncertaintyestimation, and out-of-distribution detection. However, hardware limitations ofreal-world systems constrain to smaller ensembles and lower-capacity networks,significantly deteriorating their performance and properties. We introducePacked-Ensembles (PE), a strategy to design and train lightweight structuredensembles by carefully modulating the dimension of their encoding space. Weleverage grouped convolutions to parallelize the ensemble into a single sharedbackbone and forward pass to improve training and inference speeds. PE isdesigned to operate within the memory limits of a standard neural network. Ourextensive research indicates that PE accurately preserves the properties of DE,such as diversity, and performs equally well in terms of accuracy, calibration,out-of-distribution detection, and robustness to distribution shift. We makeour code available at https://github.com/ENSTA-U2IS/torch-uncertainty.

 

Quick Read (beta)

loading the full paper ...