Notice
Recent Posts
Recent Comments
Link
| 일 | 월 | 화 | 수 | 목 | 금 | 토 |
|---|---|---|---|---|---|---|
| 1 | 2 | 3 | 4 | 5 | ||
| 6 | 7 | 8 | 9 | 10 | 11 | 12 |
| 13 | 14 | 15 | 16 | 17 | 18 | 19 |
| 20 | 21 | 22 | 23 | 24 | 25 | 26 |
| 27 | 28 | 29 | 30 | 31 |
Tags
- ENERGY-BASED MODELS FOR CONTINUAL LEARNING
- Class Incremental
- PnP algorithm
- VQ-diffusion
- 베이지안 정리
- Face Pose Estimation
- learning to prompt for continual learning
- Markov transition matrix
- CVPR2022
- CIL
- timm
- mmcv
- Facial Landmark Localization
- L2P
- Class Incremental Learning
- learning to prompt
- Energy-based model
- Img2pose
- Face Alignment
- Discrete diffusion
- img2pose: Face Alignment and Detection via 6DoF
- Continual Learning
- requires_grad
- Vector Quantized Diffusion Model for Text-to-Image Synthesis
- Mask diffusion
- state_dict()
- prompt learning
- DualPrompt
- VQ-VAE
- Mask-and-replace diffusion strategy
Archives
- Today
- Total
Computer Vision , AI
[One-page summary] A Theoretical Study on Solving Continual Learning (NeurIPS 2022) by Kim et al. 본문
Paper_review[short]
[One-page summary] A Theoretical Study on Solving Continual Learning (NeurIPS 2022) by Kim et al.
Elune001 2024. 1. 15. 20:35● Summary: Divide and conquer: Class Incremental Learning problem can be divided into within task prediction (WP) and out-of-distribution (OOD) detection and solved by optimizing each part.
● Approach highlight
- They solve CL by dividing it into TP and WP problems under the following two assumptions: 1. The domains of classes of the same task are disjoint, and 2. The domains of tasks are disjoint.
- They proved that optimizing $𝐻_{𝑇𝑃},𝐻_{𝑊𝑃}$ is equivalent to optimizing $𝐻_{𝐶𝐼𝐿}$
● Main results
● Discussion
- Using cross entropy means only capturing a subset of features and it will result in poor ODD detection because those missing features may be necessary to separate IND distribution and some out-of-distribution data. How to improve it?
'Paper_review[short]' 카테고리의 다른 글
'Paper_review[short]' Related Articles
more
