ML/DL knowledges
Collecting useful information about DL/ML
- Sigmoid VS Softmax
- Sigmoid
- probabilities produced by a Sigmoid are independent. (변수들이 독립적으로 계산됨)
- they are not constrained to sum to one. The reason for this is because the Sigmoid looks at each raw output value separately.
- Used for Binary Classification in the Logistic Regression model
- Sigmoid is equivalent to a 2 element Softmax, where the second element is assumed to be zero. Therefore, sigmoid is mostly used for binary classification.
- probabilities produced by a Sigmoid are independent. (변수들이 독립적으로 계산됨)
- Softmax
- the outputs are interrelated. The Softmax probabilities will always sum to one by design: 0.04 + 0.21 + 0.05 + 0.70 = 1.00. In this case, if we want to increase the likelihood of one class, the other has to decrease by an equal amount.
- Used for Multi-classification in the Logistics Regression model
- softmax is specially designed for multi-class and multi-label classification tasks.
- the outputs are interrelated. The Softmax probabilities will always sum to one by design: 0.04 + 0.21 + 0.05 + 0.70 = 1.00. In this case, if we want to increase the likelihood of one class, the other has to decrease by an equal amount.
- Sigmoid
-
Receptive field
CNN에서 kernel size를 통해 볼 수 있는 window 의 크기이다. receptive field 가 크면 더 넓은 영역을 cover하여 다음 feature를 뽑을 수 있고, 그 영역이 작으면 더 local 한 정보에 집중하게 된다. -
Inductive bias
-
Fourier Features
-
Over-smoothing (in graph)
GNN에서 Over-smoothing이란 Graph 층이 깊어지면서 Graph embedding이 general 한 범위에서 feature 가 지나치게 일반화되는 것을 말한다. - Maksed Auto Encoder
최근(2022년 6월 기준) 많이 사용되고 있는 self-supervised learning 방법론.
original input 의 일부를 masking 처리 없이 입력으로 줘서 encoder 를 지난 latent vector를 뽑는다.
latent vector 에 masking 처리를 하여 decoder에 입력을 주고, target image 를 예측하도록 만든다.
Reference
- https://medium.com/deep-learning-with-keras/which-activation-loss-functions-part-a-e16f5ad6d82a#:~:text=The%20practical%20reason%20is%20that,mostly%20used%20for%20binary%20classification.
- https://www.dacon.io/forum/405840