Loss Surface Simplexes for Mode Connecting Volumes and Fast Ensembling

Abstract

With a better understanding of the loss surfaces for multilayer networks, we can build more robust and accurate training procedures. Recently it was discovered that independently trained SGD solutions can be connected along one-dimensional paths of near-constant training loss. In this paper, we show that there are mode-connecting simplicial complexes that form multi-dimensional manifolds of low loss, connecting many independently trained models. Inspired by this discovery, we show how to efficiently build simplicial complexes for fast ensembling, outperforming independently trained deep ensembles in accuracy, calibration, and robustness to dataset shift. Notably, our approach only requires a few training epochs to discover a low-loss simplex, starting from a pre-trained solution. Code is available at https://github.com/g-benton/loss-surface-simplexes.

arXiv commentICML 2021

References in corpus (13)

• On Calibration of Modern Neural Networks
• Can You Trust Your Model's Uncertainty? Evaluating Predictive Uncertainty Under Dataset Shift
• The Loss Surfaces of Multilayer Networks
• Deep Ensembles: A Loss Landscape Perspective
• Qualitatively characterizing neural network optimization problems
• Essentially No Barriers in Neural Network Energy Landscape
• A Simple Baseline for Bayesian Uncertainty in Deep Learning
• Topology and Geometry of Half-Rectified Network Optimization
• Rethinking Parameter Counting in Deep Models: Effective Dimensionality Revisited
• Understanding Generalization through Visualizations
• Learning Neural Network Subspaces
• Loss Patterns of Neural Networks
• A Deep Neural Network's Loss Surface Contains Every Low-dimensional Pattern

Cited by in corpus (3)

• A Review of Uncertainty Quantification in Deep Learning: Techniques, Applications and Challenges
• Which Shortcut Cues Will DNNs Choose? A Study from the Parameter-Space Perspective
• Dangers of Bayesian Model Averaging under Covariate Shift