# Multi-Loss Weighting with Coefficient of Variations

**Authors:** Rick Groenendijk, Sezer Karaoglu, Theo Gevers, Thomas Mensink  
**arXiv:** [2009.01717](https://arxiv.org/abs/2009.01717)  
**Date:** 2020-09-03  
**Summary page:** https://www.summarizepaper.com/en/paper/2009.01717v2/

## Key points

- Optimizing an objective function defined as a weighted linear combination of multiple losses is common in machine learning and computer vision.
- Final performance of such models depends on choosing the correct relative weights for these losses.
- Traditional methods involve setting weights using an extensive grid search, which can be computationally expensive.
- The paper proposes a novel weighting scheme based on the coefficient of variations to address this issue.
- The proposed method incorporates a measure of uncertainty to balance the losses in multi-loss problems such as monocular depth estimation and semantic segmentation.
- This study focuses on single-task multi-loss problems and shows that many loss weighting methods in literature that focus on multi-task approaches do not work for those single-tasks.
- The authors demonstrate empirically that their approach is valid for depth estimation and semantic segmentation across multiple datasets.
- The proposed method sets the weights based on properties observed while training the model and evolves during training without requiring another optimization step, making it computationally efficient compared to traditional methods.

## Summary

In the field of machine learning and computer vision, optimizing an objective function that is defined as a weighted linear combination of multiple losses is a common approach for many interesting tasks. However, the final performance of such models is highly dependent on choosing the correct relative weights for these losses. Traditionally, finding a good set of weights involves adopting them into the set of hyper-parameters and setting them using an extensive grid search, which can be computationally expensive. In this paper titled "Multi-Loss Weighting with Coefficient of Variations," authors Rick Groenendijk, Sezer Karaoglu, Theo Gevers, and Thomas Mensink propose a novel weighting scheme based on the coefficient of variations to address this issue. The proposed method incorporates a measure of uncertainty to balance the losses in multi-loss problems such as monocular depth estimation and semantic segmentation. Unlike many loss weighting methods in literature that focus on multi-task approaches, this study focuses on single-task multi-loss problems and shows that these approaches do not work for those single-tasks. The authors demonstrate empirically that their approach is valid for depth estimation and semantic segmentation across multiple datasets. The proposed method sets the weights based on properties observed while training the model and evolves during training without requiring another optimization step. This eliminates the need for an extensive grid search to find optimal hyper-parameters, making it computationally efficient compared to traditional methods. Overall, this paper presents a promising solution to improve model performance by providing an effective way to choose optimal weightings for multiple losses in single-task scenarios in machine learning and computer vision applications without sacrificing computational efficiency.