Research map / Label-efficient and robust learning / Self-supervised and contrastive
Frequency-Balanced Retinal Representation Learning with Mutual Information Regularization
Seunghoon Lee, Seongjae Kang, Inhyuk Park, Gitaek Kwon, Jihyeon Baek, Doohyun Park
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 HighFreqMI regularizer (Lhmi) to MAE reconstruction (ablation) | MAE · 5 retinal benchmarks linear probing | AUROC: 0.685 -> 0.779 | Improved |
| add HighFreqMI regularizer (Lhmi) to MAE reconstruction (ablation) | MAE · 5 retinal benchmarks linear probing | AUPRC: 0.486 -> 0.614 | Improved |
| add HighFreqMI to the auxiliary-loss setting (ablation) | MAE + Laux · 5 retinal benchmarks linear probing | AUROC: 0.923 -> 0.941 | Improved |
| add HighFreqMI to the auxiliary-loss setting (ablation) | MAE + Laux · 5 retinal benchmarks linear probing | AUPRC: 0.799 -> 0.849 | Improved |
| add auxiliary feature alignment (Laux) to MAE (ablation) | MAE · 5 retinal benchmarks linear probing | AUROC: 0.685 -> 0.923 | Improved |
| add Lhmi to MAE with HF masking (ablation) | MAE w/ HF masking · 5 retinal benchmarks linear probing | AUROC: 0.679 -> 0.737 | Improved |
| add Lhmi to MAE with HF input channel (ablation) | MAE w/ HF input · 5 retinal benchmarks linear probing | AUROC: 0.746 -> 0.769 | Improved |
| feed only high-frequency tokens (25%) instead of 75% masked tokens at probing | MAE · 5 retinal benchmarks linear probing | AUROC: 0.647 -> 0.727 | Improved |
| use low-frequency tokens only instead of full input | MAE · 5 retinal benchmarks linear probing | AUROC: 0.685 -> 0.641 | Hurt |
| set lambda_hmi to 1.0 instead of 0.1 (ablation) | RetMAE (no aux) · 5 retinal benchmarks linear probing | AUROC: 0.779 -> 0.738 | Hurt |
| set lambda_hmi to 0.01 instead of 0.1 (ablation) | RetMAE (no aux) · 5 retinal benchmarks linear probing | AUROC: 0.779 -> 0.732 | Hurt |
| vary lambda_aux among 1.0, 0.1, 0.01 (ablation) | RetMAE (aux only) · 5 retinal benchmarks linear probing | AUROC: 0.927 / 0.926 / 0.932 | No clear effect |
Closest papers by meaning
- FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs (NeurIPS 2026 · similarity 0.86)
- Masked Frequency Modeling for Self-Supervised Visual Pre-Training (ICLR 2023 · similarity 0.86)
- Understanding Masked Autoencoders via Hierarchical Latent Variable Models (CVPR 2023 · similarity 0.84)
- Frequency-Guided Masking for Enhanced Vision Self-Supervised Learning (ICLR 2025 · similarity 0.83)
- Contrastive Tuning: A Little Help to Make Masked Autoencoders Forget (AAAI 2024 · similarity 0.83)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.82)
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.
