AI Paper Atlas

Research map / Label-efficient and robust learning / Recognition and long-tail

Calibrating Deep Neural Networks using Focal Loss

Jishnu Mukhoti, Viveka Kulharia, Amartya Sanyal, Stuart Golodetz, Philip Torr, Puneet Dokania

NeurIPS 2020 · Label-efficient and robust learning › Recognition and long-tail · approach: Confidence calibration objectives · 86 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
Replace cross-entropy with sample-dependent focal lossImage and text classification models · CIFAR-10, CIFAR-100, Tiny-ImageNet, 20 Newsgroups, SST BinaryECEImproved
Replace cross-entropy with focal lossCIFAR-10 classifiers · CIFAR-10 → SVHN and CIFAR-10-CAUROCImproved
Use fixed γ=3 instead of sample-dependent γ (ablation)ResNet-50 · CIFAR-10ECE (%) pre temperature scaling: 1.55 -> 1.48Hurt
Increase fixed focal-loss γ from 1 to 2 to 3 (ablation)Focal-loss classifiers · Classification datasetsECEImproved
Set focal-loss γ too high (γ≥7) (ablation)Focal-loss training · Classificationtraining successHurt
Choose γ per sample based on the correct-class probability (ablation)Focal-loss training · Multiple classification datasets and architecturescalibrationImproved
Use early stopping based on validation ECE with cross-entropy (ablation)Cross-entropy training · CIFAR-10test errorHurt
Apply temperature scaling (ablation)Focal-loss classifiers · Image and text classification datasetsECEMixed

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy