Research map / Continual learning / Online and replay-based
Where is the Truth? The Risk of Getting Confounded in a Continual World
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.
Introduces a formal framework for continual confounding and the CLEVR-based ConCon dataset, showing that continual learners (including cumulative training) get confounded, especially in the "strict" variant where joint training succeeds.
Problem
Spurious correlations (confounders) that vary across tasks in continual learning can cause models to fail on unconfounded data, a challenge not addressed by standard forgetting-focused continual learning or static confounding studies.
Approach
Formalizes continual confounding via logical constraints on the learned predicate across tasks with disjoint or strict confounder specifications, and shows that some confounded task sequences make cumulative training fail even where joint training succeeds (insidious continual confounding); operationalized in the new CLEVR-based ConCon dataset.
Key results
- On ConCon strict, NN joint training reaches 95.7% on unconfounded data while cumulative training reaches 72.6%, a gap of over 20 points.
- On ConCon disjoint, all CL methods with memory (ER, DER, BGS, FROMP, PNNs) retain near 100% on old tasks yet stay near 50% on unconfounded data, as does joint training for NN and NeSy.
- In the ImageNet preliminary experiment, cumulative training reaches 89.3% versus 91.7% for joint training on the same data.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train on all tasks jointly instead of cumulatively | ResNet-18 · ConCon strict unconfounded test | accuracy: 72.6 -> 95.7 | Helped |
| Shuffle task data so confounders are not aligned with tasks | ResNet-18 · ConCon strict unconfounded test | accuracy: 72.6 -> 89.5 | Helped |
| Train on all tasks jointly instead of cumulatively | NeSy · ConCon strict unconfounded test | accuracy: 69.9 -> 70.8 | No clear effect |
| Use cumulative training instead of joint training | ResNet-18 · ConCon disjoint unconfounded test | accuracy: 50.0 -> 49.9 | No clear effect |
| Use cumulative training instead of joint training | NeSy · ConCon disjoint unconfounded test | accuracy: 50.7 -> 50.8 | No clear effect |
| Use memory-based ER instead of naive sequential training | ResNet-18 · ConCon disjoint old task T1@T3 | accuracy: 50.0 -> 100.0 | Helped |
| Use memory-based ER instead of naive sequential training | ResNet-18 · ConCon disjoint unconfounded test | accuracy: 48.1 -> 50.1 | No clear effect |
| Train on all tasks jointly instead of cumulatively | ResNet-18 · ImageNet selected classes | accuracy: 89.3% -> 91.7% | Helped |
Surprising findings
- Even cumulative training with an infinite memory buffer fails on unconfounded data in the strict setting while joint training on the same data succeeds (insidious continual confounding).
- Preventing catastrophic forgetting does not help: CL methods that retain all old-task accuracy still fail on unconfounded data in the disjoint setting.
Limitations
- Dataset is synthetic CLEVR-based with simple single-object confounders and four objects per image
- Memory buffer fixed at 100 for memory-based methods, limiting comparisons; hyperparameters lightly tuned
- Only image modality; text and time series not covered
Experimental setup
Datasets: ConCon (built on CLEVR) CLEVR ImageNet (preliminary experiment)
Benchmarks: ConCon disjoint ConCon strict
Metrics: accuracy on current task accuracy on old tasks accuracy on unconfounded test set
Baselines: Naive Sequential EWC FROMP ER DER BGS GDUMB PNNs Shuffled Cumulative Joint
Builds on: CLEVR Experience Replay EWC FROMP DER
Closest papers by meaning
- A Unified and General Framework for Continual Learning (ICLR 2024 · similarity 0.90)
- Continual Learning with Global Alignment (NeurIPS 2024 · similarity 0.88)
- Confounder-Free Continual Learning via Recursive Feature Normalization (ICML 2025 · similarity 0.88)
- STAR: Stability-Inducing Weight Perturbation for Continual Learning (ICLR 2025 · similarity 0.87)
- Model Sensitivity Aware Continual Learning (NeurIPS 2024 · similarity 0.87)
- Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning (NeurIPS 2025 · 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.