Research map / Continual learning / Domain and prompt-based
Continual Learning in the Frequency Domain
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 CLFD, which maps inputs to the wavelet (frequency) domain via DWT and selects class-specific frequency features via similarity-based dropout, plugging into rehearsal-based continual learning methods to improve accuracy and training efficiency on edge devices.
Problem
Rehearsal-based continual learning is computationally and memory expensive, limiting deployment on resource-constrained edge devices, and naive downsampling of inputs loses information while prior frequency-domain methods (DCT, learnable DWT layers) either remove spatial information or add cross-tas
Approach
Apply a Haar discrete wavelet transform to inputs (halving spatial size while keeping spatial structure for augmentation), freeze the encoder after task 1, and use class-aware frequency dropout/semantic dropout guided by cosine similarity of class-level low-frequency features to balance reuse and interference of features across tasks, with no new cross-task learnable parameters.
Key results
- CLFD increases Class-IL accuracy of the SOTA CL method by up to 6.83% and reduces training time by 2.6x on Jetson Orin NX.
- On Split CIFAR-10 with buffer 50, CLFD-ER raises ER Class-IL accuracy from 29.42% to 45.56% while peak memory drops from 497MB to 205MB.
- CLFD-CLS-ER reaches 59.98% Class-IL on Split CIFAR-10 with buffer 125 and 18.73% on Split Tiny-ImageNet, the best results reported.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Integrate CLFD with rehearsal-based methods | ER, DER++, ER-ACE, and CLS-ER · S-CIFAR-10 and S-Tiny-ImageNet, Class-IL | accuracy | Helped |
| Integrate CLFD with rehearsal-based methods | ER, DER++, ER-ACE, and CLS-ER · S-CIFAR-10 and S-Tiny-ImageNet, Task-IL | accuracy | Mixed |
| Integrate CLFD into rehearsal methods | Rehearsal-based methods · S-CIFAR-10 and S-Tiny-ImageNet | peak memory | Helped |
| Add FFE and CFFS (ablation) | ER · S-CIFAR-10, Class-IL, buffer 50 | accuracy: 29.42 ±3.53 -> 45.56 ±3.71 | Helped |
| Add FFE (ablation) | ER without CFFS · S-CIFAR-10, Class-IL, buffer 50 | accuracy: 29.42 ±3.53 -> 39.19 ±0.83 | Helped |
| Add FFE (ablation) | ER without CFFS · S-CIFAR-10, Task-IL, buffer 50 | accuracy: 86.36 ±1.43 -> 88.01 ±0.06 | Helped |
| Add CFFS (ablation) | ER without FFE · S-CIFAR-10, Class-IL, buffer 50 | accuracy: 29.42 ±3.53 -> 37.80 ±5.78 | Helped |
| Add CFFS (ablation) | ER without FFE · S-CIFAR-10, Task-IL, buffer 50 | accuracy: 86.36 ±1.43 -> 85.78 ±2.43 | Hurt |
| Increase feature selection proportion from 10% to 90% (ablation) | CLFD-ER · S-CIFAR-10, Class-IL, buffer 50 | accuracy: 51.03 -> 39.58 | Hurt |
| Increase feature selection proportion from 10% to 90% (ablation) | CLFD-ER · S-CIFAR-10, Task-IL, buffer 50 | accuracy: 85.67 -> 87.97 | Helped |
| Increase feature selection proportion from 10% to 90% (ablation) | CLFD-ER-ACE · S-CIFAR-10, Class-IL, buffer 50 | accuracy: 50.37 -> 54.20 | Helped |
| Increase feature selection proportion from 10% to 90% (ablation) | CLFD-ER-ACE · S-CIFAR-10, Task-IL, buffer 50 | accuracy: 85.06 -> 89.83 | Helped |
Surprising findings
- Using only high-frequency wavelet components (X_lh, X_hl, X_hh) gives accuracy close to low-frequency only in some settings, and the combined FFE is best, though the reported single-component table values are partially missing in the text.
- Within CFFS, higher feature selection proportions improve Task-IL but lower proportions help Class-IL for high-plasticity methods like ER.
Limitations
- Focused only on rehearsal-based methods; applicability to rehearsal-free settings is unverified.
- Analysis limited to finite buffer sizes, not infinite-buffer scenarios.
- Trade-off: improved stability can reduce plasticity on new tasks.
Experimental setup
Datasets: Split CIFAR-10 Split Tiny-ImageNet ImageNet-R
Benchmarks: Split CIFAR-10 Split Tiny-ImageNet Split ImageNet-R
Metrics: Class-IL accuracy Task-IL accuracy peak memory footprint training time training FLOPs forgetting stability-plasticity trade-off
Baselines: oEWC SI LwF ER DER++ ER-ACE CLS-ER SparCL SGD JOINT
Builds on: ER DER++ ER-ACE CLS-ER SparCL Mammoth
Closest papers by meaning
- Enhancing Visual Continual Learning with Language-Guided Supervision (CVPR 2024 · similarity 0.87)
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.87)
- LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual Learning (ICLR 2025 · similarity 0.87)
- Continual Learning Using a Kernel-Based Method Over Foundation Models (AAAI 2025 · similarity 0.87)
- Re-evaluating Continual Learning with Few-Shot Adaptation (NeurIPS 2026 · similarity 0.86)
- CLDyB: Towards Dynamic Benchmarking for Continual Learning with Pre-trained Models (ICLR 2025 · similarity 0.85)
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.