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 |
- pointcloud
- sensorfusion
- imageprocessing
- differentiablerendering
- realtimerendering
- Slam
- turtlebot
- alphablending
- vectorcalculus
- LIDAR
- Ros
- usbcamera
- rospackages
- covariancematrix
- adaptivedensitycontrol
- 3dgaussiansplatting
- tilebasedrasterizer
- turtlebot3
- electromagnetism
- rostopics
- gaussiansplatting
- roslaunch
- vectorfields
- opencv
- rosnoetic
- raspberrypi
- 3dmapping
- catkinworkspace
- NERF
- ComputerVision
- Today
- Total
Wiredwisdom
Deep-learning point 본문
`R2=1-(SST)/(SSE)=1-(Sigma(y_text(target)-bar y_text(predict))^2)/ (Sigma(y_text(target)-bar y_text(average))^2)`
R2 > 0.8: 매우 좋은 모델
R2 0.6~0.8: 괜찮은 모델
R2 0.4~0.6: 개선 여지가 있는 모델
R2 < 0.4: 추가적인 개선이 필요한 모델
입력 데이터의 Normalization의 이점
1. 그래디언트 폭주를 막을 수 있다.(Gradient Explosion)
발산하게 되는 경우
2. Chain rule 영향
x = 입력값
w = 가중치
z = w * x # 첫 번째 계산
y = sigmoid(z) # 활성화 함수 적용
L = (target - y)² # 손실 함수
`(delL)/(delw)=(delL)/(dely)*(dely)/(delz)*(delz)/(delw)`
∂L/∂y = -2(target - y) # 손실함수의 미분
∂y/∂z = y * (1-y) # 시그모이드 함수의 미분
∂z/∂w = x # 선형 함수의 미분
`(delL)/(delw) = -2(target - y) * y(1-y) * x`
여기서 tartget =1 , y(predict) = 0.5 라 가정할때
# 입력이 작은 경우
x = 2 gradient = -2 * (0.5) * 0.2 * 2 = -0.4 # 안정적인 값
# 입력이 큰 경우
x = 1000 gradient = -2 * (0.5) * 0.2 * 1000 = -200 # 매우 큰 값
3. 그래디언트 포화(Saturation)
시그모이드 함수같이 극대값이 1이나 0으로 수렴값이 존재하는 경우
입력값이 크면 값이 포화되어 학습저하가 발생된다.
'Artificial Intelligence > Deep Learning Basic' 카테고리의 다른 글
Why MSE (0) | 2025.03.30 |
---|---|
Cross-entropy and MSE principles (0) | 2025.03.30 |
Neural Network Basic (0) | 2024.08.01 |
Attention (0) | 2024.06.23 |
RNN - LSTM - LLM summary (0) | 2024.06.21 |