| 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 |
- Markov transition matrix
- Discrete diffusion
- DualPrompt
- img2pose: Face Alignment and Detection via 6DoF
- CIL
- VQ-VAE
- ENERGY-BASED MODELS FOR CONTINUAL LEARNING
- Mask-and-replace diffusion strategy
- PnP algorithm
- Vector Quantized Diffusion Model for Text-to-Image Synthesis
- learning to prompt
- Facial Landmark Localization
- Face Alignment
- requires_grad
- state_dict()
- learning to prompt for continual learning
- Face Pose Estimation
- L2P
- Continual Learning
- CVPR2022
- Img2pose
- prompt learning
- Mask diffusion
- Class Incremental
- timm
- 베이지안 정리
- Class Incremental Learning
- mmcv
- VQ-diffusion
- Energy-based model
- Today
- Total
목록CIL (3)
Computer Vision , AI
[One-page summary] ENERGY-BASED MODELS FOR CONTINUAL LEARNING (CoLLAs 2022) by Li et al
● 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 cos..
[One-page summary] DualPrompt: Complementary Prompting for Rehearsal-free Continual Learning (ECCV 2022) By Wang et al.
●Summary: Learning two disjoint prompt spaces makes rehearsal-free prompt-based continual learning more effectively ●Approach highlight ○Using task-invariant General prompt(L_g) at ViT 1~2nd layer and task-specific Expert Prompt(# of task ×L_e) at ViT 3~5th layer ○Prefix tuning: before MSA, concatenate key prompt to hidden representation of key and value prompt to hidden representation of value,..
[One-page summary] Learning to Prompt for Continual Learning (CVPR 2022) By Wang et al.
Category: continual learning (class incremental) ●Summary: Use prompt with a pre-trained model for rehearsal buffer-free and task-agnostic continual learning ●Approach highlight ○Prompt pool memory space allows rehearsal buffer-free and task-agnostic ○ Penalize frequently-used prompts by using prompt frequency table H_t at training time for the diversity of prompt (1) ○ At training-time, If quer..