Research map / Continual learning / Continual fine-tuning of LLMs
LANCE: Low Rank Activation Compression for Efficient On-Device Continual Learning
Marco Apolinario, Kaushik Roy
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 |
|---|---|---|---|
| fine-tune 4 layers with LANCE instead of 2 layers (ablation) | LANCE · ImageNet-1k Split B | accuracy: 59.18 -> 62.20 | Improved |
| replace full BP with LANCE (4 layers) | ResNet18 · ImageNet-1k Split B | accuracy / memory: 64.28 -> 62.20 accuracy, 61.25 -> 16.59 MB | Mixed |
| replace full BP with LANCE (2 layers) | ResNet34 · CIFAR-100 (last 2 layers) | accuracy: 76.19 -> 74.89 | Hurt |
| replace iterative HOSVD with one-shot LANCE | ResNet18 · CIFAR-100 (last 4 layers) | accuracy: 75.35 -> 75.7 | No clear effect |
| replace ASI with LANCE | ResNet18 · CIFAR-100 (last 4 layers) | accuracy: 71.93 -> 75.7 | Improved |
| replace ASI with LANCE | MobileNetV2 · CUB-200 (last 4 layers) | accuracy: 50.69 -> 61.21 | Improved |
| replace iterative HOSVD with LANCE | MCUNet · Raspberry Pi 3B+ latency | training time: 2.1x-4.4x slower -> LANCE lowest | Improved |
| replace ASI with LANCE | MCUNet · Raspberry Pi 3B+ latency | training time: 19-76% slower -> LANCE lowest | Improved |
| replace full BP with LANCE | MCUNet · Raspberry Pi 3B+ latency | training time: BP 3-9% slower | Improved |
| raise energy threshold epsilon (ablation) | LANCE · CUB-200 | accuracy / memory | Mixed |
| increase number of calibration batches N (ablation) | LANCE · CIFAR-100 | accuracy | No clear effect |
| lower energy threshold epsilon (stronger compression) (ablation) | LANCE · gradient angle vs full BP | angle: within ~70 degrees at up to two orders of magnitude compression | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 2 on most shared results.
- CODE-CL — CODE-CL: Conceptor-Based Gradient Projection for Deep Continual Learning (ICCV 2025) · better on 0 of 2 shared results
- TRGP — TRGP: Trust Region Gradient Projection for Continual Learning (ICLR 2022) · better on 0 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
- Beyond Low-rank Decomposition: A Shortcut Approach for Efficient On-Device Learning (ICML 2025 · similarity 0.87)
- SparCL: Sparse Continual Learning on the Edge (NeurIPS 2022 · similarity 0.86)
- PLAN: Proactive Low-Rank Allocation for Continual Learning (ICCV 2025 · similarity 0.85)
- PaLoRA: Paced Low-Rank Adaptation for Continual Learning (NeurIPS 2026 · similarity 0.85)
- TinyTrain: Resource-Aware Task-Adaptive Sparse Training of DNNs at the Data-Scarce Edge (ICML 2024 · similarity 0.84)
- C-LoRA: Continual Low-Rank Adaptation for Pre-trained Visual Models (NeurIPS 2026 · similarity 0.83)
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.
