Research map / Label-efficient and robust learning / Self-supervised and contrastive
Hallucination Improves the Performance of Unsupervised Visual Representation Learning
Jing Wu, Jennifer Hobbs, Naira Hovakimyan
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 Hallucinator to MoCoV2 | MoCoV2 (ResNet-18) · CIFAR-10 linear eval | top-1 accuracy: 87.21 -> 89.23 | Improved |
| add Hallucinator to SimSiam | SimSiam (ResNet-18) · STL-10 linear eval | top-1 accuracy: 87.79 -> 88.16 | Improved |
| add Hallucinator to MoCoV1 | MoCoV1 (ResNet-50) · ImageNet-1K linear eval (100 epochs) | top-1 accuracy: 57.27 -> 59.17 | Improved |
| add Hallucinator to SimCLR | SimCLR (ResNet-50) · ImageNet-1K linear eval (100 epochs) | top-1 accuracy: 61.23 -> 61.71 | Improved |
| add all modules (center cropping, extrapolation, hallucination) to MoCoV2 (ablation) | MoCoV2 (ResNet-50) · ImageNet-200 linear eval | top-1 accuracy: 62.57 -> 63.58 | Improved |
| add center cropping alone (ablation) | MoCoV2 (ResNet-50) · ImageNet-200 linear eval | top-1 accuracy: 62.57 -> 62.58 | No clear effect |
| add asymmetric extrapolation alone (ablation) | MoCoV2 (ResNet-50) · ImageNet-200 linear eval | top-1 accuracy: 62.57 -> 62.82 | Improved |
| add non-linear hallucination layers (n=2 or 3) instead of identity n=0 (ablation) | MoCoV2 (ResNet-50) · ImageNet-200 linear eval | top-1 accuracy: 62.85 -> 63.58 | Improved |
| reduce center-crop ratio p below 0.5 (ablation) | MoCoV2 (ResNet-50) · ImageNet-200 linear eval | top-1 accuracy: 63.58 -> 63.21 | Hurt |
| use negative extrapolation range (-0.5, 0.0) (ablation) | MoCoV2 (ResNet-50) · ImageNet-200 linear eval | top-1 accuracy: 63.58 -> 62.80 | Hurt |
| add Hallucinator to MoCoV2 pretraining for COCO transfer | MoCoV2 · COCO instance segmentation | AP^mk: 34.2 -> 34.6 | Improved |
Closest papers by meaning
- Hallucination-aware Intermediate Representation Editing in Large Vision-Lanugage Models (ICLR 2026 · similarity 0.88)
- Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding (CVPR 2024 · similarity 0.87)
- SECOND: Mitigating Perceptual Hallucination in Vision-Language Models via Selective and Contrastive Decoding (ICML 2025 · similarity 0.87)
- Mitigating Object Hallucination in Large Vision-Language Models via False Discovery Controlled Visual Data Splitting (NeurIPS 2026 · similarity 0.87)
- Decoupling Contrastive Decoding: Robust Hallucination Mitigation in Multimodal Large Language Models (NeurIPS 2025 · similarity 0.87)
- Learning from Fine-Grained Visual Discrepancies: Mitigating Multimodal Hallucinations via In-Context Visual Contrastive Optimization (ICML 2026 · similarity 0.86)
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.
