Research map / Continual learning / Class-incremental
ESSENTIAL: Episodic and Semantic Memory Integration for Video Class-Incremental Learning
Jongseo Lee, Kyungho Bae, Kyle Min, Gyeong-Moon Park, Jinwoo Choi
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 |
|---|---|---|---|
| add MR module and semantic memory (global) to episodic-memory-only baseline (ablation) | ESSENTIAL · SSV2 (10x9) | average incremental accuracy (%): 42.1 -> 47.4 | Improved |
| make MR module and semantic memory task-specific instead of global (ablation) | ESSENTIAL · SSV2 (10x9) | average incremental accuracy (%): 47.4 -> 48.9 | Improved |
| use class-specific rather than task-specific semantic memory (ablation) | ESSENTIAL · SSV2 (10x9) | average incremental accuracy (%): 48.9 -> 47.9 | Hurt |
| use cross-attention with semantic prompt instead of MLP, interpolation, add or multiply (ablation) | MR module · SSV2 (10x9) | average incremental accuracy (%): 45.8 -> 48.9 | Improved |
| add both static and temporal matching losses (ablation) | ESSENTIAL · SSV2 (10x9) | average incremental accuracy (%): 46.3 -> 48.9 | Improved |
| add only one matching loss (static or temporal) (ablation) | ESSENTIAL · SSV2 (10x9) | average incremental accuracy (%): 46.3 -> 46.9 (static), 47.1 (temporal) | Improved |
| increase semantic prompt length from 8 to 24 without static matching (ablation) | ESSENTIAL · SSV2 (10x9) | average incremental accuracy (%): 47.1 -> 48.2 (below 48.9 for length 8 with static matching) | No clear effect |
| train MR module with sparse instead of dense input features (ablation) | ESSENTIAL · SSV2 (10x9) | average incremental accuracy (%): 47.6 -> 48.9 | Improved |
| uniform frame sampling instead of random or temporal-attention sampling (ablation) | ESSENTIAL · SSV2 (10x9) | average incremental accuracy (%): 48.2 -> 48.9 | Improved |
| use MR module and semantic prompt at inference (ablation) | ESSENTIAL (global) · SSV2 (10x9) | average incremental accuracy (%): 47.4 -> 47.8 | No clear effect |
| use learned semantic prompt instead of random prompt for retrieval (ablation) | MR module · SSV2 (feature L2 distance) | L2 distance to dense features: 33.2 -> 8.2 | Improved |
| use cross-attention instead of self-attention in temporal encoder (ablation) | ESSENTIAL · SSV2 (10x9) | average incremental accuracy (%): 48.1 -> 48.9 | Improved |
Closest papers by meaning
- SEC-Prompt:SEmantic Complementary Prompting for Few-Shot Class-Incremental Learning (CVPR 2025 · similarity 0.87)
- Bisecle: Binding and Separation in Continual Learning for Video Language Understanding (NeurIPS 2025 · similarity 0.87)
- Pre-trained Vision and Language Transformers Are Few-Shot Incremental Learners (CVPR 2024 · similarity 0.87)
- ViLCo-Bench: VIdeo Language COntinual learning Benchmark (NeurIPS 2024 · similarity 0.86)
- Learning Conditional Space-Time Prompt Distributions for Video Class-Incremental Learning (CVPR 2025 · similarity 0.86)
- Long-Tail Class Incremental Learning via Independent Sub-prototype Construction (CVPR 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.
