티스토리 뷰
목적 : 이것도 AAnet Introduction 에서 existing CIL methods로 소개했다. 빨리 읽고 12시이후부터는 코딩해봐야한다.
Mnemonics Training: Multi-Class Incremental Learning without Forgetting
Abstract
- Multi-Class Incremental Learning(MCIL) 목표 :이전 개념에 대해 trained model을 incrementally 갱신하여 새로운 개념을 학습시키는 것
- inherent trade-off : Catastrophic forgetting vs learning new concepts
- representativeness of these examples에 크게 좌우된다. 본 논문에서는 mnemonics라는 novel and automactic framework를 제시하는데 examplars를 parametize하고 그것을 end-to-end 방식으로 최적화 할 수 있도록 한다.
- bi-level optimization을 통해 프레임워크를 훈련하고 세가지 실험(CIFAR100,ImageNet-Subset, ImageNet)에서 실험하니 mnemonics exemplars를 쓰는 것이 최첨단 기법보다 큰 폭으로 성능을 향상시켰다.
- 신기하게도 mnemonics exemplars는 boundaries beween classes 근처에 자주 위치한다. (신기하네)
- Random : exemplars들이 early phase에 무작위적으로 샘플링 됨.
- Herding : exemplars들이 early phase에 mean sample의 최근접 이웃들임.
'AI' 카테고리의 다른 글
Gaussian Mixture Model 의 Clustering (0) | 2024.06.24 |
---|---|
abs summray for searching prototype (0) | 2024.06.15 |
[LUCIR] CVPR'2019 (1) | 2024.06.01 |
Herding Dynamical Weights to Learn (0) | 2024.06.01 |
Adaptive Aggregation Networks for Class-Incremental Learning (0) | 2024.05.28 |
공지사항
최근에 올라온 글
최근에 달린 댓글
- Total
- Today
- Yesterday
링크
TAG
- /var/log/fontconfig.log
- ctfloader
- torchmodel
- Arbitrary image stylization
- Deformable Part Model
- neural network applications
- batch란
- TransferLearning
- 1-stage detection
- /var/log/faillog
- DetectoRS
- Group Acitivty Recognition
- tf2to1
- Deformable Convolutional Networks
- 딥페이크탐지
- /var/log/tallylog
- tensorflow convert
- 도메인 어댑테이션
- max-margin_loss
- textual guidance
- Chatgpt4
- alternatives.log
- StyleCLIP
- 첫논문
- 샴네트워크
- /var/log/alternatives.log
- Source-free unsupervised domain adaptation
- yolov3 #tensorflow_convert
- tf version porting
- 전이학습
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 |
글 보관함