Research map / Label-efficient and robust learning / Recognition and long-tail
Bias Loss for Mobile Neural Networks
Lusine Abrahamyan, Valentin Ziatchin, Yiming Chen, Nikos Deligiannis
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 changed | Where | Measured | Effect |
|---|---|---|---|
| add bias loss (ablation) | MobileNetV3 1.0x · ImageNet val | top-1 accuracy: 75.2 -> 75.7 | Improved |
| add skip blocks (ablation) | MobileNetV3 1.0x (CE loss) · ImageNet val | top-1 accuracy: 75.2 -> 75.5 | Improved |
| add bias loss to SkipblockNet-M (ablation) | SkipblockNet-M · ImageNet val | top-1 accuracy: 75.5 -> 76.2 | Improved |
| add skip blocks instead of widening the network (ablation) | MobileNetV3 · ImageNet val | top-1 accuracy: 75.3 -> 75.5 | Improved |
| train with bias loss instead of cross-entropy | ShuffleNetV2 0.5x · CIFAR-100 | top-1 accuracy: 69.5 -> 71 | Improved |
| train with bias loss instead of cross-entropy | SqueezeNet · CIFAR-100 | top-1 accuracy: 69.4 -> 70.4 | Improved |
| train with bias loss instead of cross-entropy | MobileNetV2 0.75x · CIFAR-100 | top-1 accuracy: 68 -> 68.6 | Improved |
| train with bias loss instead of focal loss | DenseNet (k=12) · CIFAR-100 | top-1 accuracy: 79.5 -> 79.9 | Improved |
| replace MobileNetV3 Large with SkipblockNet-M | mobile classifier · ImageNet val | top-1 accuracy: 75.2 -> 76.2 | Improved |
| replace MobileNetV2 with SkipblockNet-S as SSD backbone | SSD · PASCAL VOC detection | mAP: 73.1 -> 73.6 | Improved |
| train with bias loss instead of cross-entropy (ablation) | SkipblockNet-M · last conv layer feature variance | average variance: 0.09 -> 0.15 | Improved |
| use SkipblockNet-M instead of MobileNetV3 1.0x (both CE) | feature extractor · last conv layer feature variance | average variance: 0.05 -> 0.09 | Improved |
Closest papers by meaning
- MobileODE: An Extra Lightweight Network (NeurIPS 2025 · similarity 0.86)
- MnasNet: Platform-Aware Neural Architecture Search for Mobile (CVPR 2019 · similarity 0.86)
- CNN Filter DB: An Empirical Investigation of Trained Convolutional Filters (CVPR 2022 · similarity 0.85)
- ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices (CVPR 2018 · similarity 0.85)
- Large-Margin Softmax Loss for Convolutional Neural Networks (ICML 2016 · similarity 0.85)
- Beyond Filters: Compact Feature Map for Portable Deep Model (ICML 2017 · similarity 0.85)
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.
