Research map / Continual learning / Class-incremental
Quantum-Gated Task-interaction Knowledge Distillation for Pre-trained Model-based Class-Incremental Learning
Linjie Li, Huiyu Xiao, Jiarui Cao, Zhenyu Wu, Yang Ji
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 quantum gating module (QGTM) (ablation) | adapter-based PTM CIL baseline · CIFAR100 B0-Inc20 | average incremental accuracy: 78.42 -> 88.59 | Improved |
| add quantum gating module (QGTM) (ablation) | adapter-based PTM CIL baseline · CUB200 B0-Inc20 | average incremental accuracy: 65.26 -> 91.73 | Improved |
| add quantum gating module (QGTM) (ablation) | adapter-based PTM CIL baseline · ImageNet-A B0-Inc10 | final accuracy: 37.89 -> 51.79 | Improved |
| add task-interaction distillation loss L_QKD (ablation) | QKD with QGTM · CIFAR100 B0-Inc20 | average incremental accuracy: 88.59 -> 92.69 | Improved |
| add task-interaction distillation loss L_QKD (ablation) | QKD with QGTM · ImageNet-A B0-Inc10 | final accuracy: 51.79 -> 54.98 | Improved |
| add sparsity regularization L_s (ablation) | QKD with QGTM and L_QKD · CIFAR100 B0-Inc20 | average incremental accuracy: 92.69 -> 93.76 | Improved |
| add sparsity regularization L_s (ablation) | QKD with QGTM and L_QKD · CUB200 B0-Inc20 | final accuracy: 89.42 -> 90.20 | Improved |
| use quantum gate instead of MLP task recognizer (ablation) | QKD · ImageNet-R B0-Inc20 | average accuracy: 80.25 -> 81.79 | Improved |
| use quantum gate instead of cosine similarity recognizer (ablation) | QKD · ImageNet-R B0-Inc20 | average accuracy: 80.32 -> 81.79 | Improved |
| use quantum gate instead of attention recognizer (ablation) | QKD · CUB200 B0-Inc20 | average accuracy: 92.57 -> 93.54 | Improved |
| feed first-task adapter A1 features to the gate (ablation) | QKD · CIFAR100 B0-Inc10 | final accuracy: 88.83 -> 90.20 | Improved |
| use two adapters [A1,A2] instead of A1 as gate input (ablation) | QKD · CIFAR100 B0-Inc10 | final accuracy: 90.20 -> 89.74 | Hurt |
Closest papers by meaning
- Disentangling Representations through Multi-task Learning (ICLR 2025 · similarity 0.83)
- QuanTA: Efficient High-Rank Fine-Tuning of LLMs with Quantum-Informed Tensor Adaptation (NeurIPS 2024 · similarity 0.83)
- TAID: Temporally Adaptive Interpolated Distillation for Efficient Knowledge Transfer in Language Models (ICLR 2025 · similarity 0.82)
- Disentangling and mitigating the impact of task similarity for continual learning (NeurIPS 2024 · similarity 0.82)
- Class Incremental Learning via Likelihood Ratio Based Task Prediction (ICLR 2024 · similarity 0.82)
- Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · 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.
