Research map / Continual learning / Online and replay-based
Memory-Integrated Reconfigurable Adapters: A Unified Framework for Settings with Multiple Tasks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MIRA, which stores LoRA-style adapters as values in a Hopfield-style associative memory and learns retrieval keys so a frozen ViT-B/16 CLIP backbone can recall task-specific adapter mixtures across domain generalization, domain-incremental, and class-incremental learning.
Problem
Existing continual and generalization methods are studied in isolation, and parameter-efficient adapters lack explicit memory mechanisms for retrieving task-specific knowledge at inference across DG, DIL, and CIL settings.
Approach
Store weight adapters as values in an associative memory (Hopfield network) and learn the retrieval keys post hoc, so per-sample affine combinations of adapters are retrieved to modulate a frozen backbone, rather than using fixed keys or storing raw data.
Key results
- MIRA reports 82.73% CIL/DIL average accuracy on iDigits, 88.64 on CORe50, and 68.24 on DomainNet, outperforming ICON (next best) in average accuracy on each.
- On DG benchmarks (PACS, VLCS, OfficeHome, DomainNet), MIRA averages 81.92 versus 80.48 for GESTUR and 80.30 for PEGO.
- Ablation: adapters per task increasing from 1 to 10 raises average accuracy from 64.68 to 65.89; affine separation function gives the best average (65.89).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MIRA instead of ICON | iDigits CIL · iDigits CIL | Avg. Acc.: 71.53$\pm$0.68 -> 83.00$\pm$1.29 | Helped |
| Use MIRA instead of ICON | iDigits DIL · iDigits DIL | Avg. Acc.: 84.83$\pm$0.51 -> 82.46$\pm$0.12 | Hurt |
| Use MIRA instead of ICON | CORe50 CIL · CORe50 CIL | Avg. Acc.: 80.85$\pm$0.23 -> 83.39$\pm$0.24 | Helped |
| Use MIRA instead of ICON | CORe50 DIL · CORe50 DIL | Avg. Acc.: 89.01$\pm$0.33 -> 93.89$\pm$0.33 | Helped |
| Use MIRA instead of ICON | DomainNet CIL · DomainNet CIL | Avg. Acc.: 65.43$\pm$0.15 -> 67.29$\pm$0.19 | Helped |
| Use MIRA instead of ICON | DomainNet DIL · DomainNet DIL | Avg. Acc.: 54.44$\pm$0.21 -> 69.18$\pm$0.10 | Helped |
| Replace affine separation with Softmax (ablation) | MIRA · CIL | accuracy: 67.29 -> 66.87 | Hurt |
| Replace affine separation with ReLU (ablation) | MIRA · CIL | accuracy: 67.29 -> 66.60 | Hurt |
| Replace affine separation with Tanh (ablation) | MIRA · CIL | accuracy: 67.29 -> 66.73 | Hurt |
| Increase adapter count from 1 to 10 (ablation) | MIRA · CIL | accuracy: 63.75 -> 67.29 | Helped |
| Increase adapter count from 1 to 10 (ablation) | MIRA · DIL | accuracy: 69.08 -> 69.18 | Mixed |
| Increase adapter count from 1 to 10 (ablation) | MIRA · DG | accuracy: 61.21 -> 61.19 | No clear effect |
Limitations
- Improvements over the strongest baseline in DG on some datasets are small or comparable (e.g., VLCS where PEGO is best).
- Gains from more adapters saturate beyond about 5 adapters per task.
Experimental setup
Datasets: iDigits CORe50 DomainNet PACS VLCS OfficeHome
Benchmarks: iDigits CORe50 DomainNet PACS VLCS OfficeHome
Metrics: average accuracy forgetting
Baselines: Fine-tuning EWC LwF L2P S-Prompts DualPrompt CODA-P LAE ICON SWAD CLIP SMA
Builds on: PEGO LoRA CLIP Universal Hopfield Networks L2P
Closest papers by meaning
- MemSkill: Learning and Evolving Memory Skills for Self-Evolving Agents (NeurIPS 2026 · similarity 0.86)
- Rapid Learning without Catastrophic Forgetting in the Morris Water Maze (ICML 2024 · similarity 0.86)
- MemGen: Weaving Generative Latent Memory for Self-Evolving Agents (ICLR 2026 · similarity 0.85)
- MemEvolve: Meta-Evolution of Agent Memory Systems (ICML 2026 · similarity 0.85)
- Mem-T: Densifying Rewards for Long-Horizon Memory Agents (ICML 2026 · similarity 0.85)
- Learning to Continually Learn via Meta-learning Agentic Memory Designs (NeurIPS 2026 · similarity 0.84)
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.