PolyLoss : A new framework for loss functions

Original Paper: [ https://arxiv.org/pdf/2204.12511.pdf]

Authors: Zhaoqi Leng, Mingxing Tan , Chenxi Liu , Ekin Dogus Cubuk , Xiaojie Shi, Shuyang Cheng ,Dragomir Anguelov [Waymo And Google]


Authors propose a new framework of loss functions, motivated by the Taylor series expansion of commonly used functions like cross entropy and focal loss.

When CE and focal losses are expanded, it’s easier to see the similarities and differences between them. This seems to be the prime motivation to create a framework of losses which express such common functions as a special case.

The authors also experiment with the coefficients of taylor expansion of these functions and assess their impact on training a ResNet-50 model on ImageNet-1k dataset. Based on their experiments, altering the coefficients of the first N terms of the expansion series seems to give the best results.

Upon further simplification, L(poly-1) seems to provide superior result as compared to cross entropy loss on a variety of different tasks involving detection, classification and segmentation.

There’s a lot to love about the simplicity and impact of this paper. A one line change in the loss function seems to give better results than the traditionally used loss functions.

Originally published at https://clearsignal.xyz on May 18, 2022.



Computer vision engineer; excited about building products and companies; clearsignal.xyz

Love podcasts or audiobooks? Learn on the go with our new app.

Get the Medium app

A button that says 'Download on the App Store', and if clicked it will lead you to the iOS App store
A button that says 'Get it on, Google Play', and if clicked it will lead you to the Google Play store