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 |
Tags
- 일상
- momentum #adagrad #adam #early stopping #data augmentation #dropout #label smoothing #batch normalization
- flops
- recall
- 대학원생
- 부트캠프
- map
- confusion matrix
- deep learning #machine learning #history
- 네이버
- neural networks #linear #nonlinear #perceptrons
- Optimization
- generalization #overfitting # bootstrapping #bagging #boosting #stochastic gradient desent #mini-batch gradient descent #batch gradient descent
- CNN #padding #stride #parameters
- FPS
- 1x1 convolution
- evaluation
- precision
- objectdetection
- AI tech
- IOU
Archives
- Today
- Total
공부 일기장
Historical Review 본문
Deep learner?
1. Implemenation Skills
2. Math Skills
3. Knowing a lot of recent Papers
Key components of Deep learning
(아래와 같은 항목에 focus 해서 논문을 살펴보는 것이 좋다)
- Data
- Model
- Loss fuction
- Optimization(Algorithm)
1. Data
data depend on the type of the problem to solve
2. Model
model의 성질에 따라 다양한 결과가 도출됨
3. Loss
Loss fuction is a proxy
4. Optimization(Algorithm)
- dropout
- early stopping
- k-fold validation
- weight decay
- batch normalizaion
- mix up
- ensemble
Deep Learning History
논문 - Deep Learning's Most Important Ideas - A Breif Historical Review
https://dennybritz.com/blog/deep-learning-most-important-ideas
2012 - AlexNet
- 이 시기를 기준으로 deeplearning 모델들이 실제적으로 성능을 발휘
- 모든 기계학습의 판도가 바뀜
2013 - DQN
- deepmind (알파고)
- 강화학습
2014 - Encoder / Decoder
- 번역 task
- A를 잘 encoding해서 B로 decoding
2015 - Adam Optimizer
- 다양한 optimizer가 있는데 adam이 결과가 일반적으로 잘 나옴
2016 - GAN, ResNet
- 생성모델 GAN
- 네트워크를 깊게 쌓는 ResNet
- 기존의 layer가 깊어지면 training이 안되는것이 아니라 test성능이 떨어짐
- 많이 쌓아도 test data 성능이 나오도록
2017 - Transfomer
- Attention is all you need
2018 - Bert
- bert자체의 중요성도 있지만 finetuned nlp models의 개념이 탄생한 것이 더 중요
2019 - Big Language Models(GPT)
- OpenAI
- fine-tuned nlp models의 끝판왕
- 굉장히 많은 파라미터
2020 - Self-supervised Learning
- 2020년도의 트렌드
- SimCLR
- 한정된 데이터를 사용하여 Task를 수행하는 것이 일반적이였다면 unlabeled 데이터를 가지고 학습하여 수행
- 좋은 visual representation을 학습하는 것이 중요하다.
'네이버 부스트 캠프 > DL basic' 카테고리의 다른 글
ILSVRC(modern models) (0) | 2023.03.06 |
---|---|
Convolutional Neural Networks (0) | 2023.02.27 |
Optimization (0) | 2023.02.27 |
Neural Networks & Multi-Layer Perceptron (0) | 2023.02.27 |
Comments