Research map / Continual learning / Domain and prompt-based
Sequential Mastery of Multiple Visual Tasks: Networks Naturally Learn to Learn and Forget to Forget
Guy Davidson, Michael C. Mozer
Check 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 |
|---|---|---|---|
| Learn additional related tasks before introducing the next task | Standard CNN, sequential training · Synthetic visual tasks, episodes 3–10 | trials to criterion | Improved |
| Repeatedly retrain earlier tasks across episodes | Standard CNN, sequential training · Synthetic visual tasks, episodes 1–10 | trials to criterion | Improved |
| Add task-specific learned biases to the first convolutional layer (ablation) | Standard CNN, sequential training · Synthetic visual tasks, especially episodes 2–3 | trials to criterion | Improved |
| Add task-specific learned biases to the first convolutional layer (ablation) | Standard CNN, sequential training · Synthetic visual tasks, later episodes | trials to criterion | No clear effect |
| Mix tasks from color, shape, and texture dimensions instead of using one dimension (ablation) | Standard CNN, sequential training · Synthetic visual tasks, episodes 1–10 | trials to criterion | No clear effect |
| Mix tasks from different dimensions (ablation) | Standard CNN, sequential training · Synthetic visual tasks, first two episodes | backward interference | Improved |
| Apply MAML-style meta-training, splitting examples between task learning and meta-training (ablation) | Standard CNN, sequential training · Synthetic visual tasks, episodes 1–10 | trials to criterion | Hurt |
| Use alternative ratio or power curricula rather than the balanced old/new split (ablation) | Standard CNN, sequential training · Synthetic visual tasks, episodes 1–10 | trials to criterion | No clear effect |
| Train on all ten tasks simultaneously instead of introducing tasks sequentially (ablation) | Standard CNN · Color and texture tasks | trials to criterion | Hurt |
| Train on all ten tasks simultaneously instead of introducing tasks sequentially (ablation) | Standard CNN · Shape tasks | trials to criterion | No clear effect |
| Fit an exponential rather than power-law forgetting curve (ablation) | Forgetting-curve model · Synthetic visual tasks, held-out forgetting curves | held-out MSE: 0.0232 -> 0.0192 | Improved |
| Fit an exponential rather than power-law forgetting curve (ablation) | Forgetting-curve model · Synthetic visual tasks, training forgetting curves | training MSE: 0.0045 -> 0.0198 | Hurt |
Closest papers by meaning
- Continual Learning in the Teacher-Student Setup: Impact of Task Similarity (ICML 2021 · similarity 0.90)
- A continual learning survey: Defying forgetting in classification tasks (TPAMI 2021 · similarity 0.89)
- Separating the 'what' and 'how' of compositional computation to enable reuse and continual learning (NeurIPS 2025 · similarity 0.87)
- Optimal Task Order for Continual Learning of Multiple Tasks (ICML 2025 · similarity 0.87)
- Mixture of Experts Meets Prompt-Based Continual Learning (NeurIPS 2024 · similarity 0.87)
- Continual Learning with Global Alignment (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.
