Research map / Continual learning / Continual fine-tuning of LLMs
Incremental Task Learning with Incremental Rank Updates
Rakib Hyder, Ken Shao, Boyu Hou, Panos Markopoulos, Ashley Prater-Bennette, M. Salman Asif
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 |
|---|---|---|---|
| raise initial rank from 1 to 6, keeping increment 1 (ablation) | low-rank ITL · P-MNIST | average accuracy (%): 74.23 -> 82.21 | Improved |
| raise initial rank from 6 to 11, keeping increment 1 (ablation) | low-rank ITL · P-MNIST | average accuracy (%): 82.21 -> 85.61 | Improved |
| raise initial rank from 1 to 11, keeping increment 1 (ablation) | low-rank ITL · R-MNIST | average accuracy (%): 81.57 -> 91.09 | Improved |
| raise rank increment per task from 1 to 2 (initial rank 11) (ablation) | low-rank ITL · P-MNIST | average accuracy (%): 85.61 -> 90.51 | Improved |
| raise rank increment per task from 2 to 4 (initial rank 11) (ablation) | low-rank ITL · P-MNIST | average accuracy (%): 90.51 -> 93.84 | Improved |
| raise rank increment per task from 1 to 4 (initial rank 11) (ablation) | low-rank ITL · R-MNIST | average accuracy (%): 91.09 -> 94.12 | Improved |
| raise rank increment per task from 1 to 4 (initial rank 11) (ablation) | low-rank ITL · parameter count (P/R-MNIST) | # parameters: 0.11M -> 0.2M | Hurt |
| update only the last 1 layer vs last 5 layers of ResNet18 (ablation) | low-rank ITL · S-CIFAR100 (ResNet18) | average accuracy (%): 34.38 -> 65.03 | Improved |
| use rank-1 ITL instead of Orthog Subspace | incremental task learning · S-miniImageNet (3-layer MLP) | average accuracy (%): 38.1 -> 54.7 | Improved |
| use rank-1 ITL instead of HAT | incremental task learning · P-MNIST (3-layer MLP) | average accuracy (%): 90.1 -> 85.6 | Hurt |
| use rank-1 ITL instead of Adam-NSCL | incremental task learning · S-CIFAR-100 (ResNet18) | average accuracy (%): 74.31 -> 68.46 | Hurt |
| use rank-1 ITL instead of Parallel 2 (similar parameter count) | incremental task learning · S-CIFAR100 (3-layer MLP) | average accuracy (%): 62.8 -> 65.9 | 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.
- ICARL — iCaRL: Incremental Classifier and Representation Learning (CVPR 2017) · better on 2 of 2 shared results
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
- Class Incremental Learning via Likelihood Ratio Based Task Prediction (ICLR 2024 · similarity 0.89)
- GACL: Exemplar-Free Generalized Analytic Continual Learning (NeurIPS 2024 · similarity 0.87)
- Learn more, but bother less: parameter efficient continual learning (NeurIPS 2024 · similarity 0.87)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.87)
- Replay-and-Forget-Free Graph Class-Incremental Learning: A Task Profiling and Prompting Approach (NeurIPS 2024 · similarity 0.86)
- Continual Learning with Global Alignment (NeurIPS 2024 · 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.
