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 |
- Homomorphic Encryption
- GPT
- bert
- matrix multiplication
- Pre-training
- Transformer
- Knowledge Tracing
- math
- Deep learning
- Knowledge Distillation
- attention
- Github Copilot
- 딥러닝
- ICML
- 표현론
- Private ML
- AI
- Data Augmentation
- Model Compression
- Computer Vision
- KT
- 자연어처리
- NLP
- Machine Learning
- Copilot
- Language Modeling
- Residual Connection
- 머신러닝
- 동형암호
- Natural Language Processing
- Today
- Total
목록Machine Learning & Deep Learning/Computer Vision (2)
Anti Math Math Club
![](http://i1.daumcdn.net/thumb/C150x150.fwebp.q85/?fname=https://blog.kakaocdn.net/dn/crDGBr/btqI5SSWy7L/dWnJvoF4uJ3Kcu6N0laVXk/img.png)
이번 포스팅에서는 ICLR 2020 Honorable Mention Award를 받은 Generative Pretraining from Pixels라는 논문에 대해서 알아보도록 하겠습니다. Computer Vision에서는 ImageNet을 이용한 pre-training & fine-tuning이 거의 표준으로 자리잡고 있는 반면, NLP에서는 Wikipedia같은 거대한 corpus가 있다고 해도 sentiment나 POS tag같은 label이 붙어 있는 데이터는 많지 않기 때문에 pre-training을 하는 것이 쉽지 않습니다. 하지만, GPT와 BERT를 기점으로 한 self-supervised learning은 이러한 데이터를 활용해서 모델을 똑똑하게 pre-train하는 것을 가능하게 해 주..
![](http://i1.daumcdn.net/thumb/C150x150.fwebp.q85/?fname=https://blog.kakaocdn.net/dn/diU0SX/btqI4TxZ8jV/7AxO31YR3mp8sMRUwbt3uk/img.png)
이번 포스트에서는 (제가 알기로는) Consistency Regularization이 처음 소개된 논문인 Regularization With Stochastic Transformations and Perturbations for Deep Semi-Supervised Learning이라는 논문에 대해서 알아봅시다. Consistency Regularization이란, 간단히 말해서 모델의 Input에 augmentation을 가해서 새로운 input을 만들었을 때, output (prediction)이 별로 변하지 않아야 한다는 가정을 바탕으로 모델을 regularize하는 방법 입니다. 예를 들어서, 이미지를 분류하는 CNN이 하나 있을 때, 기존에 있던 강아지 사진을 뒤집거나 돌리는 등의 작업을 해서 ..