| 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 |
- CVPR2022
- state_dict()
- VQ-VAE
- L2P
- Continual Learning
- Discrete diffusion
- Markov transition matrix
- img2pose: Face Alignment and Detection via 6DoF
- Face Alignment
- ENERGY-BASED MODELS FOR CONTINUAL LEARNING
- learning to prompt for continual learning
- mmcv
- learning to prompt
- Mask diffusion
- prompt learning
- Img2pose
- PnP algorithm
- requires_grad
- Vector Quantized Diffusion Model for Text-to-Image Synthesis
- 베이지안 정리
- Mask-and-replace diffusion strategy
- timm
- Energy-based model
- Face Pose Estimation
- Class Incremental Learning
- CIL
- Facial Landmark Localization
- DualPrompt
- VQ-diffusion
- Class Incremental
- Today
- Total
Computer Vision , AI
[One-page summary] ENERGY-BASED MODELS FOR CONTINUAL LEARNING (CoLLAs 2022) by Li et al 본문
[One-page summary] ENERGY-BASED MODELS FOR CONTINUAL LEARNING (CoLLAs 2022) by Li et al
Elune001 2023. 4. 18. 18:44● Summary: Adopts energy-based method models classifier to continual learning to solve class incremental problem
●Approach highlight
○ Energy-based model for classifier: take class y and data x and as input and output is their energy value
○Energy-based models loss function makes minimize energy of positive pair and maximize energy of negative pair
● Main Results
● Discussion
○ Computational cost: computational cost of prediction is proportional to the number of labels
○ Still has low performance compared to iCaRL with a replay and still low performance when increasing the number of tasks
EBM classifier: y(판별하고 싶은 class들)과 x(입력)를 입력으로 받아 Energy-based classifier에 넣어서 가장 낮은 에너지 준위가 나오는 y가 해당 x의 클래스
