Research map / Continual learning / Class-incremental
Class Incremental Learning via Likelihood Ratio Based Task Prediction
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 TPL, which predicts task-id in class-incremental learning via a likelihood ratio between each task's distribution and its complement (estimated from replay data), combined with a logit score on top of HAT-based task models with a pre-trained DeiT.
Problem
Task-id prediction is the key bottleneck in class-incremental learning (CIL) when no task identifier is given at test time; existing TIL+OOD approaches use traditional OOD detectors that do not exploit CIL-specific information.
Approach
Formulate task-id prediction as a binary hypothesis test between task t and its complement, using a likelihood ratio where the in-task density is estimated by Mahalanobis distance and the complement density by KNN over replay buffer features, plus a maximum-logit score combined via an energy-based OR composition, with each task model trained as an OOD-style classifier with an extra "others" class
Key results
- TPL achieves average Last ACC of 76.21% and AIA of 83.23% across five settings with pre-trained DeiT-S/16, outperforming ROW (73.72% Last) and PASS (68.25% Last).
- Without pre-training, TPL reaches 57.5% average Last ACC, versus 54.2% for DER, the next best.
- With full-ImageNet pre-training (TPL_PFI), average Last ACC is 85.22% and AIA 89.42%, with near-zero forgetting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace vanilla HAT_CIL task-id scoring with the likelihood-ratio score (ablation) | HAT_CIL · Five datasets (average) | Last accuracy (%): 63.41\% -> 71.25\% | Helped |
| Use MLS OOD scoring alone instead of vanilla HAT_CIL scoring (ablation) | HAT_CIL · Five datasets (average) | Last accuracy (%): 63.41\% -> 68.69\% | Helped |
| Combine the likelihood-ratio and MLS scores (ablation) | HAT_CIL · Five datasets (average) | Last accuracy (%): 71.25\% -> 76.21\% | Helped |
| Replace the MLS logit score with EBO (ablation) | TPL · Five datasets (average) | Last accuracy (%): 76.21\% vs. 75.76\% | No clear effect |
| Replace the MLS logit score with MSP (ablation) | TPL · Five datasets (average) | Last accuracy (%): 76.21\% -> 71.32\% | Hurt |
| Use smaller replay buffers (ablation) | TPL · Five datasets (average) | Last accuracy (%): 76.21 -> 75.56 | Hurt |
| Use DeiT pretrained with full ImageNet rather than the 611-class ImageNet model | TPL · Five datasets (average) | Last accuracy (%): 76.21 -> 85.22 | Helped |
Surprising findings
- Traditional forgetting-rate formula is argued to be inappropriate for CIL because accuracy drops due to more classes rather than forgetting.
- KNN, the best OOD method in prior work, performs poorly as a task-id scorer in CIL when limited to replay data.
Limitations
- Relies on replay buffer samples to estimate the complement distribution, so performance depends on buffer size (smaller buffers degrade results somewhat).
- Uses pre-trained DeiT-S/16 with a reduced ImageNet subset (611 classes) for main results to avoid information leak.
Experimental setup
Datasets: CIFAR-10 CIFAR-100 TinyImageNet ImageNet (611 classes for pre-training)
Benchmarks: CIFAR-10 5-task CIFAR-100 10-task CIFAR-100 20-task TinyImageNet 5-task TinyImageNet 10-task
Metrics: Last accuracy average incremental accuracy (AIA) forgetting rate
Baselines: iCaRL A-GEM EEIL GD DER++ HAL DER FOSTER AFC BEEF MORE ROW
Builds on: HAT CSI Mahalanobis distance KNN OOD detection MLS energy-based models
Closest papers by meaning
- H2ST: Hierarchical Two-Sample Tests for Continual Out-of-Distribution Detection (CVPR 2025 · similarity 0.90)
- eTag: Class-Incremental Learning via Embedding Distillation and Task-Oriented Generation (AAAI 2024 · similarity 0.89)
- Learning Task-Aware Language-Image Representation for Class-Incremental Object Detection (AAAI 2024 · similarity 0.89)
- Replay-and-Forget-Free Graph Class-Incremental Learning: A Task Profiling and Prompting Approach (NeurIPS 2024 · similarity 0.88)
- Defying Imbalanced Forgetting in Class Incremental Learning (AAAI 2024 · similarity 0.87)
- GACL: Exemplar-Free Generalized Analytic Continual Learning (NeurIPS 2024 · similarity 0.87)
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.