Research map / Continual learning / Online and replay-based
FlyPrompt: Brain-Inspired Random-Expanded Routing with Temporal-Ensemble Experts for General Continual Learning
Hongwei Yan, Guanglong Sun, Kanglei Zhou, Qian Li, Liyuan Wang, Yi Zhong
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 REAR analytic random-expanded router to prompt experts (ablation) | FlyPrompt · CIFAR-100 Si-Blurry (Sup-21K) | A_auc: 80.75 -> 81.90 (no EMA); 82.17 -> 83.24 (with EMA) | Improved |
| add REAR router (ablation) | FlyPrompt (Sup-21K/1K) · CIFAR-100 | A_auc: 72.51 -> 78.48 | Improved |
| add EMA temporal-ensemble heads (ablation) | FlyPrompt · CIFAR-100 Si-Blurry (Sup-21K) | A_auc: 81.90 -> 83.24 | Improved |
| add EMA heads to single-prompt baseline without REAR (ablation) | single prompt · CIFAR-100 Si-Blurry (Sup-21K) | A_auc: 71.33 -> 71.69 | No clear effect |
| add prompt experts (multi-prompt across tasks) (ablation) | FlyPrompt · CIFAR-100 Si-Blurry (Sup-21K) | A_auc: 71.33 -> 80.75 | Improved |
| replace expert routing with RanPAC analytic classifier (ablation) | FlyPrompt · CIFAR-100 Si-Blurry (Sup-21K) | A_auc: 83.24 -> 69.91 | Hurt |
| plug REAR into DualPrompt (ablation) | DualPrompt · CIFAR-100 (Sup-21K) | A_auc: 76.04 -> 80.63 | Improved |
| plug REAR into MVP (ablation) | MVP · CIFAR-100 (Sup-21K) | A_auc: 67.74 -> 67.44 | No clear effect |
| plug TE2 into MVP (ablation) | MVP · CIFAR-100 (Sup-21K) | A_auc: 67.74 -> 68.91 | Improved |
| use two EMA heads (0.9, 0.99) instead of online head only (ablation) | FlyPrompt · CIFAR-100 | A_auc: 81.90 -> 83.24 | Improved |
| use a single slow EMA head 0.999 (ablation) | FlyPrompt · CIFAR-100 | A_auc: 81.90 -> 81.80 | No clear effect |
| use SoftMax + element-wise max instead of mean ensemble (ablation) | FlyPrompt · CIFAR-100 | A_auc: 81.34 -> 83.24 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- DualPrompt — DualPrompt: Complementary Prompting for Rehearsal-Free Continual Learning (ECCV 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- A Unified Continual Learning Framework with General Parameter-Efficient Tuning (ICCV 2023 · similarity 0.87)
- Mixture of Experts Meets Prompt-Based Continual Learning (NeurIPS 2024 · similarity 0.87)
- Fly-CL: A Fly-Inspired Framework for Enhancing Efficient Decorrelation and Reduced Training Time in Pre-trained Model-based Continual Representation Learning (ICLR 2026 · similarity 0.87)
- BiLoRA: Almost-Orthogonal Parameter Spaces for Continual Learning (CVPR 2025 · similarity 0.87)
- InTAct: Interval-based Task Activation Consolidation for Continual Learning (NeurIPS 2026 · similarity 0.86)
- FlyLoRA: Boosting Task Decoupling and Parameter Efficiency via Implicit Rank-Wise Mixture-of-Experts (NeurIPS 2025 · 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.
